Discover Applied Sciences

Review

Automated diagnosis and classification of liver cancers using deep

learning techniques: a systematic review
Sarthak Grover1 · Surbhi Gupta2

Received: 8 August 2024 / Accepted: 17 September 2024

© The Author(s) 2024  OPEN

Abstract
Liver cancer is one of the main causes of cancer-related mortality globally. It is a rising threat with over 700,000 deaths
and 800,000 new cases annually. To tackle this, deep learning (DL) artificial intelligence (AI) tools have been researched
extensively to create predictive models that aid in the early diagnosis and classification of liver cancers. The liver, however,
is prone to metastasis from various gastrointestinal cancers, making it difficult to correctly classify tumors present in the
organ. This paper is a thorough systematic evaluation of several published studies that used deep learning algorithms
to diagnose, classify, and predict common liver tumors, including but not limited to Hepatocellular carcinoma and
Intrahepatic cholangiocarcinoma. Articles selected in this review have been published between 2016 and 2024 from
Google Scholar, Research Gate, Science Direct, IEEE Xplore, Springer Link, and other resources as per the PRISMA guide-
lines. Amongst these, the highest recorded accuracy was 99.38%, with other studies showing close results at 97% and
below. The highest recorded sensitivity was 100%. Selected studies were also analyzed based on their choice of dataset,
DL algorithm, and preprocessing techniques. Magnetic resonance imaging and computed tomography (CT) are by far
the preferred medium for imaging data because of their high resolution; however, CT is used more often because it’s
cheaper and provides a better view of extrahepatic space. Genetic datasets showed great potential in differentiating
between primary and secondary liver cancers. Nonetheless, several gaps were identified across the current literature. The
problem of accurately differentiating cancer types in the liver still persists. However, the use of 3D imaging datasets aids
in this significantly. Though, it is clear that 3D datasets are underutilized. Dynamic imaging is also beneficial and should
be considered more often. For future research, we strongly recommend the use of Gaussian Mixture Model, Watershed
Transform, Sparsity classification, DLIR and Siamese algorithms as they each showed promising results. Lastly, we have
shown a benchmark framework for a predictive AI model that can be referred to develop future models. Though numer-
ous past methods have produced excellent prediction results, the prevalence of liver cancer-related deaths has not yet
been reduced. To address the difficulties in the field of cancer prediction, more thorough research is required, along with
the development of more efficient decision support systems using deep learning.

Article Highlights

• The article provides a comprehensive evaluation of various diagnostic approaches used for liver cancers and their
potential in Deep Learning diagnostic models.
• We present a benchmark framework for an AI-based computational model for the diagnosis of diseases
• The article presents novel approaches for preprocessing and diagnostic techniques for liver cancer diagnosis

* Surbhi Gupta, [email protected]; Sarthak Grover, [email protected] | 1Oakridge International School, Bangalore,
Karnataka, India. 2Model Institute of Engineering and Technology, Jammu, J&K, India.

Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

Keywords Liver cancer · Artificial intelligence · AI-based prediction · Diagnosis · Decision support system · Deep learning

1 Introduction

Cancer is a genetic disease caused by changes in the genes that control cell function, especially cell growth and division [1]. It
can be inherited, caused by damage to DNA, or simply mutations and errors occurring during cell division processes. Cancer
formation is described as uncontrolled cell division, where the body’s cells continue to proliferate, ignoring signals, such as cyc-
lins, to manage the rate [2]. All tumors and patients have a different combination of genetic factors and mutations influencing
the growth of their tumor, making cancer difficult to fight [2, 3]; hence, proactive measures such as early diagnosis and predic-
tion are vital and can save numerous lives. As of 2022, WHO claimed there were 20 million new cases of cancer worldwide [4].
The diagnosis and classification of various types of cancer become difficult because of metastasis and the presence of tissue
changes that are not cancer. Metastasis is the process by which cancer cells break away from the original (primary) tumor and
spread via blood, lymphatic system, etc., to new locations in the body where they form new (secondary) tumors [5]. Additionally,
Hyperplasia (Hypergenesis) is another process that can make cancer diagnosis challenging. Hyperplasia occurs when cells in a
tissue multiply rapidly, causing extra cells to build up [6]. While Hyperplasia may be a precursor to cancer or other abnormalities,
it can also be a build-up of normal cells, making diagnosis difficult and confusing [7]. Similarly, dysplastic cells (abnormal cells),
which are usually not cancerous, appear anomalous when observed. This can cause problems when diagnosing cancer through
imaging techniques. One example of this is the difficulty that occurs when differentiating between dysplastic nodules and early
Hepatocellular Carcinoma in the liver [8].
Liver cancer is one of the most prevalent cancer variants, regarded as the 6th most-diagnosed cancer in 2020 [9]. New
research shows that more than 800,000 people are diagnosed with liver cancer each year [10]. With this, as of 2020, liver
cancer accounted for 8.3% of total cancer deaths [11]. Liver cancer symptoms also present late, meaning patients show-
ing symptoms are likely to have serious cases [12]. The most common forms of primary liver cancers are Hepatocellular
Carcinoma (HCC) and Intrahepatic Cholangiocarcinoma (ICC) [13]. However, because of the liver’s dual blood supply,
it is highly susceptible to secondary tumors (metastasis) from various other gastrointestinal cancers [14]. Furthermore,
the sinusoidal epithelium, which comprises cells that make up the walls of the hepatic sinusoids (specialized capillaries),
contains fenestrations that enable malignant cells to penetrate the walls and enter the liver parenchyma [15, 16]. Metas-
tases also occur because of the connection between the blood supply of the intestines and the liver via the portal vein. To
support this, research showed that around 70% of patients with colorectal cancers eventually develop liver metastases as
well [17]. The prevalence of metastasized tumors in the liver makes the classification of liver tumors extremely difficult.
Liver cancer as a threat is also growing. Figure 1 shows the persistent increase in liver cancer death rate in the USA
over a decade.
Research published in the Journal of Hepatology predicted that the number of new liver cancer cases annually would
increase by around 50% by 2040—meaning 1.4 million new yearly cases. This emphasizes the need for efficient diagnostic
systems that can intake large amounts of data, and thus handle more patients more quickly [18].
Alpha-fetoprotein tumor marker tests, liver function tests, and biopsies are all common, yet invasive, diagnosis methods for
liver cancers [19]. Imaging, however, is also a viable tool. Research on artificial intelligence-deep learning (AI-DL) models for the
automated diagnosis and classification of cancers using imaging like CT or MRI has been rising. There have been great strides
in the use of AI for medical purposes, like faster diagnosis, improved accuracies, specificity, and sensitivities comparable to or

Fig. 1  Liver cancer deaths in

the USA per year from 2008 to
2017 [95]

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

outperforming expert radiologists in the respective fields. Thus, AI-based models have significant potential for clinical applica-
tion for cancer diagnosis. Computer-aided diagnosis provides higher than before accuracy and also enables minimal manual
interaction. This also makes these tools applicable for use in areas with few medical professionals. With this, this review aims to
highlight the progression of deep learning in the diagnosis and classification of liver cancers. Various novel models are analyzed
to understand what is working and what is not. Integrative models have also been analyzed, where additional modifiers have
been incorporated to improve feature extraction, liver lesion segmentation, delineation, etc. Both of which are highly useful for
tumor classification. The advantages of the datasets and models used are also explored. This paper is a comprehensive review
intended to highlight the contribution of different papers in developing techniques for the purpose of tackling liver cancer
diagnosis and to understand what can be improved.

1.1 Contributions of manuscript

This study comprises an exhaustive systematic review of various published studies utilizing deep learning algorithms to
diagnose, classify and prognose common liver cancers. Also, we provide reasoning behind the complications of diagnos-
ing liver cancers, with an emphasis on metastases. Our primary research contributions are as follows:

• A detailed understanding of liver cancer detection methods utilizing conventional diagnostic techniques is examined
in the study. And, this study provides comparisons between popularly used diagnostic procedures for liver cancers.
This is done to help decide and approach the dataset to use for future such computational models.
• The study proposes an efficient AI-based Computational framework for the diagnosis of cancer.
• The review highlights various gaps identified among the studies analyzed, identifying the scopes for improvement.
The challenges associated with detecting liver cancer through deep learning are emphasized. The future scope of
current studies has also been noted.
• Suggestions for future computational models are also presented, backed with evidence extracted from the analysis
of existing works.

1.2 Organization of manuscript

Section 2: Here we show the research process used for this review, following the PRISMA guidelines. The criteria used to
select the studies are shown, and the research investigations being answered are also listed.
Section 3: This section details the background knowledge, and explains various popular diagnosis methodologies used
around the globe. We primarily discuss methods that are used to train predictive AI models. Additionally, this section also
details the structure and function of popular AI Deep Learning algorithms that are discussed later. We have also made an
attempt to explain the workings of the most commonly used DL algorithms, namely ANN, CNN, FCN, and SVM algorithms.
Section 4: The literature survey section conducts a comparative analysis between diagnostic imaging tools used for
training computational predictive models, and elaborates use of 2D and 3D datasets for the same. Lastly, this section
provides the literature review, where existing studies on liver cancer predictive models are systematically analyzed based
on their dataset used, objective, challenges, and future scope.
Section 5: In this section, we show a benchmark framework for an AI-based predictive model that can be built on to
create future models.
Section 6: This section provides solutions to all the research investigations we mentioned earlier. We state the evalu-
ation parameters used for AI models and provide descriptions of popular liver cancer datasets of many types. This sec-
tion also discusses the research gaps that have been identified through the literature review, and we present detailed,
evidence-backed suggestions for future computational models and areas for research.
Section 7: The conclusion section summarizes the future potential of the use of Deep Learning in the field of disease
diagnosis, and in terms of liver cancers.

2 Materials and methods

This review has been conducted in accordance with the PRISMA guidelines. The literature review was conducted during
the months of January and February 2024, using Google Scholar, Research Gate, Science Direct, IEEE Xplore, etc. as the
sources for the works. Each of these sources are search engines filled with academic research only. For the papers gained

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

from manual searches, keywords such as these were used: “Liver Cancer Diagnosis”, “Deep Learning”, “Machine Learning”,
“Liver Cancer Classification”, “Artificial Intelligence” and “HCC and ICC diagnosis”. Overall, original research papers with
the objective of using artificial intelligence ML/DL models for the classification, diagnosis, and prognosis of various liver
cancers were incorporated in this review. The exhaustive selection criteria followed to attain all the research articles can
be seen in Table 1. These criteria were based on the PRISMA flowchart given in Fig. 2.
Table 1 below shows the inclusion and exclusion criteria that were used to select the studies for this review. The PRISMA
Flow Diagram has been based on these criteria.

2.1 Research investigations

This study aims to provide solutions for 4 research investigations. Below are the 4 investigations:

• Research Investigation 1: What are the evaluation parameters for AI ML/DL predictive models?
• Research Investigation 2: List commonly available datasets for liver tumors
• Research Investigation 3: What research gaps have been identified in recent studies?
• Research Investigation 4: What potential research opportunities can be suggested?

3 Background study

This section provides an overview of various diagnostic techniques used for liver cancer, including the most commonly
utilized imaging (MRI, CT and US) and invasive techniques (miRNA and DNA methylation). Also, the section digs into
methods hospitals use to maximize the efficacy of these tools through processes like dynamic or contrast-enhanced imag-
ing. We then discuss the most common and efficient Deep Learning algorithms implemented for predictive AI models.

3.1 Diagnostic procedures commonly used for liver cancers

This section provides descriptions of various diagnostic methods used for liver cancers, specifically those that are com-
monly used as datasets for training Artificial Intelligence models.

3.1.1 Imaging techniques

• MRI (Magnetic resonance imaging): MRI is a non-invasive imaging technique used to take detailed images of organs,
bones, muscles, and blood vessels. This is done by passing current to coiled wires, to create an electrical field in the
patient’s body [20]. Radiowaves are then sent through the body to create a detailed image of internal structures. When
time and money are not an issue, MRI is usually the preferred choice over CT when soft tissue or organs need to be
viewed. This is because MRI has been shown to be a better indicator of differences between normal and abnormal
tissue [21].
• CT (Computerized tomography): CT scanning is another commonly used non-invasive screening technique. It is done
through the use of X-ray waves to create an image of the body. CT scans use machines that rotate rapidly in circles
around the patient so that X-ray beams are shot at many different angles, providing greater detail and making it
superior to X-ray scans by themselves [22]. Once scanned, CT data is then cut into cross-sectional 2D images called
slices (also done in MRI), and this is what specialists review to make a medical decision. [23] It is commonly used to
detect tumors and internal bleeding/damage [22].
• Ultrasound imaging: Ultrasound scans (non-invasive) use high-frequency sound waves to create images of internal
structures in the body [24]. The scan is performed through an ultrasound transducer. Ultrasound waves are sent into
the body and move through it. The waves that reflect back are detected by the transducer, and used to produce an
image of the structures inside the body [25]. Liver ultrasound has many benefits; it can show signs of fat storage in
your liver (steatotic liver disease), hepatitis, fibrosis or cirrhosis, and liver lesions [26].

Vol:.(1234567890)
 Discover Applied Sciences
(2024) 6:508

Table 1  Inclusion and exclusion parameters

S. no Parameters Inclusion standards Exclusion standards

1 Period Original academic research conducted between 2016 and 2024 Original academic research conducted before 2016
2 Investigations Academic research focused on the diagnosis of different liver cancers Academic research focused on diseases other than liver cancers
3 Comparator Academic research works aiming to predict the diagnosis, prognosis, or Academic research works proposing predictive models other than for liver cancers
classification of liver cancers
4 Methodology Academic research articles using ML/DL methods Academic research articles using methods other than ML/DL
5 Design of Study Original research works comprising experimental results Review papers, case studies, or papers in languages other than English
| https://doi.org/10.1007/s42452-024-06218-0

Vol.:(0123456789)
Review
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

Fig. 2  PRISMA flow diagram

3.1.1.1 Use of contrast agents Contrast agents are commonly used in MRI and CT screenings. A contrast is a substance
injected into the body, orally or through an IV [22]. The use of contrast-enhanced MRI (CE-MRI) or contrast-enhanced
CT (CECT) is common when diagnosing liver masses because the contrast improves the delineation of the suspected
region in the patient, making the reports easier to review. Contrast agents can have side effects on the patient and are
therefore used only when necessary. In our context, contrast-enhanced CT is recommended for detecting masses and
malignancies [27].

3.1.1.2 Dynamic imaging Dynamic imaging is used with both CT and MRI screenings. Dynamic imaging is performed by
taking sequential images (CT or MRI) over a period of time to understand the intricacies of tissue microvasculature. It can
be used to capture changes in the body over time [28]. Take CT for example: The functionality of regular CT and dynamic
CT is identical, except that dynamic CT involves taking a series of series of images continuously over time. This allows for
monitoring of internal processes like blood flow. Differences in blood volume and vascular permeability are evident in
dynamic imaging and can be indicative of problems, like cancers [29].

3.1.2 Alternative methods for cancer diagnosis used in AI models

Other than imaging, DNA/blood profiling is a common, but invasive, tool for diagnosis. However, profiling data is also
commonly used to train AI predictive models, thus is discussed here.

3.1.2.1 Micro RNA (miRNA) profiling for cancer diagnosis Micro RNA (miRNA) are small non-coding RNA molecules
involved in the regulation of gene expression. These molecules account for roughly 1–5% of the human genome [30],
and regulate over half of all protein-encoding genes [31]. miRNA regulate gene expression by binding to messenger
RNA (mRNA) molecules in the cell. Once bound, the mRNA will either be destroyed or recycled. Such processes reduce
the rate of translation and therefore the protein being coded for by that mRNA is underexpressed [32]. It is possible for
one miRNA molecule to bind and affect many mRNAs, and multiple miRNAs affect only one. Because of this, a single

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

miRNA is able to regulate and control various cellular processes, including proliferation, apoptosis, etc. [31]. This and
more research on miRNA concluded that numerous human diseases showed an irregular expression of miRNA [33].
Additionally, miRNA that are secreted outside of cells into extracellular fluid in exosomes, known as circulating miRNA,
have been identified as biomarkers [34] for various human diseases and even cancers [35]. miRNA profiling has shown
that miRNA expression is different in normal and tumor tissues, and can therefore be used to distinguish between them
[31]. This discovery brings a lot of potential for early diagnosis of neoplasms, allowing for proactive action and special-
ized treatment methods. The latest advances in sequencing platforms for whole-genome analysis of miRNA expression
have identified aberrantly expressed miRNA profiles corresponding to different cancers (breast, colorectal, and prostate
cancers, etc.) [31]. Such profiles are what is used to train Artificial Intelligence models to be able to identify miRNA expres-
sion profiles that correspond to a specific cancer.

3.1.2.2 DNA methylation profiling for cancer diagnosis DNA Methylation is an epigenetic tag involving the transfer of a
methyl group to the 5-carbon position of a cytosine, usually located in the promotor region of a gene [36]. Methylation
of a gene has the ability to regulate its transcription: methylation of cytosine in the promoter region of a gene inhibits
the binding of transcription factors to that gene, thus downregulating the rate of transcription. Methylation also acts
by recruiting proteins involved in gene expression, slowing down the process [36]. DNA methylation is common in cells
for development; however, abnormal DNA methylation commonly disrupts signaling mechanisms between cells, result-
ing in various diseases including cancer [37]. To understand and profile the pattern of methylation in a cell, CpG sites–
regions on DNA with a higher density of cytosine and guanine nucleotides–are analyzed [36]. More specifically, DMRs
(differentially methylated regions) are analyzed, which are multiple consecutive CpG sites and occur throughout the
genome. As mentioned before, DMRs are found mostly in the promoter regions of genes [38].
In order to utilize DNA methylation as a biomarker for disease, we look for abnormalities such as hypermethylation
and hypomethylation. Hypermethylation, an increase in methylation, of the promoter sequence results in gene silenc-
ing, whereas hypomethylation is associated with upregulation [38]. For cancers, research has linked the methylation of
promoter sequences of tumor suppressor genes to be an early sign of cancer [39]. Often, the development of cancer is
caused by the hypermethylation of tumor suppressor genes and hypomethylation of oncogenes [40]. There are various
methods to quantify methylation levels, the most popular of which is whole-genome bisulfite sequencing (WGBS) [41].
However, to understand the extent of the regulation mechanism of DNA methylation, the expression profiles of the cor-
responding mRNA can also be checked [40].
Certain genes such as RASSF1A and p16 have been shown to commonly be highly methylated in many cancers.
However, there are also other specific genes that are methylated for specific cancers, and this knowledge can be used
to differentiate between them [42]. When training AI to classify cancers or tumors, DNA methylation profiles of specific
cancer patients can be used, so that specific genes and methylation sites can be identified that are abnormally methyl-
ated and thus probably associated with that cancer. Methylation profiles of patients can also then be inputted into AI
to identify abnormalities in methylation sites and expression amount, which can then be cross-checked to existing data
on methylation patterns in different cancers.

3.2 Deep learning models

The following are some of the most popularized and efficient Deep Learning algorithms used for predictive AI. The
implementation of these algorithms is observed in Sect. 4.

3.2.1 Artificial neural networks (ANNs)

ANNs are computer network systems built of layers of code that mimic the structure and functionality of the human
nervous system. The human nervous system and neural network comprise signals being transported across neurons,
where sensory neurons send signals through the nervous system, eventually producing an output through axons to
effectors. ANN is similar to this, as it is a hierarchical network made up on many neurons, created to learn and synthe-
size new information from its input information and experiences–similar to the ability of the human brain. Without any
instructions, ANNs are able to learn from examples that they are trained on, and thus able to perform the desired task.
Their efficiency and performance depend directly on the amount of data they have been trained on. ANNs are built up
of 3 layers: input layer, hidden layer, and output layer. The number of hidden layers can vary as the architecture and
requirements of the models change [43, 44] (Fig. 3).

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

Fig. 3  Architecture of ANN

In the structure of an ANN, interconnected nodes called neurons are organized in layers. Input data is processed
through these layers of neurons, which adjust weights and identify patterns.
Features: These are labeled features in the input image. Weights: Adjustable parameters that determine the influence
and strength of each feature on the neurons in the hidden layers. Hidden Layer: These are layers of neurons that process
input features based on weights. This is how models learn complex patterns within the data.
In our context of liver cancer diagnosis, we have presented briefly the strengths and limitations of ANNs:

– Strengths: ANNs excel at learning from large datasets, making them suitable for applications requiring pattern rec-
ognition, such as liver cancer detection. ANNs can model complex, non-linear relationships, making them applicable
for many applications in liver cancer diagnosis, such as predicting cancer risk based on patient data.
– Limitations: The performance of ANNs depends heavily on the availability and quality of large datasets. In liver cancer
diagnosis, acquiring extensive labeled datasets can be challenging. ANNs are prone to overfitting, especially with
limited data, which can lead to reduced generalization and accuracy in diagnosing new, rare (unseen) cancer cases.

3.2.2 Convolutional neural networks (CNNs)

The use of CNNs has gained popularity significantly because of their impressive feature extraction. CNN is a deep learn-
ing algorithm based on the animal visual cortex and therefore is used widely for imaging-related data, such as CT and
MRI. They have been designed to automatically learn features from images. CNNs are used because, with their structure,
they are able to accurately extract relevant features from input data without the need for traditional methods like texture
analysis. They also don’t require segmentation of images for training.

Fig. 4  Convolutional Neural

Network

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

CNNs are made up of 3 layers: convolution layers, pooling layers, and fully connected layers. The first 2 layers are
involved in feature extraction, and the fully connected layer maps the extracted features out as the final output, thus
classifying the feature [43, 45]. As illustrated in the figure below, CNN architecture is used to perform object detection in
an image task. Figure 4 shows the layers and architecture of a CNN model and below we have explained the functionali-
ties of each of these layers.
Input: The image being analyzed. Convolutional + pooling layers: These layers act as feature extractors. Each convolu-
tion + pooling layer narrows down and the features selected get more precise. The more features and detailed the output
can be, the more such layers. Fully connected layers (FC): The FC layers act as classifiers. Probability values are assigned
to the features extracted, and then FC layers predict the class in which the image belongs.
In our context of liver cancer diagnosis, we have presented briefly the strengths and limitations of CNNs:

– Strengths: MRI and CT imaging are the most efficient, cheap and safe ways of diagnosing and screening liver cancers.
CNNs excel in extracting relevant features from medical imaging data without requiring manual segmentation, mak-
ing them highly accurate and crucial in this context.
– Limitations: Training and using CNNs does, however, require computational resources not readily available in all clini-
cal settings. This makes them difficult to implement. The decision-making process of CNNs can be opaque, making
it difficult for clinicians to interpret the results and understand the rationale behind certain diagnoses. This is vital
information, especially when diagnosing life-threatening diseases like cancer, as it can help in prognosis and patient
well-being.

3.2.3 Fully convolutional networks (FCNs)

FCNs are a modification of a CNN, in which fully connected layers are converted into convolutional layers, and decon-
volutional layers are added in order to attain an output map from the original image [46]. FCNs are majorly used for
segmentation, or more specifically, semantic segmentation. Semantic segmentation is a deep learning algorithm in
which every pixel in an input image is assigned a label [47]. This way, the algorithm forms collections of pixels within the
whole image, isolating clearly all separate features. FCNs are extremely beneficial for feature extraction for this reason.
In our context of liver cancer diagnosis, we have presented briefly the strengths and limitations of FCNs:

– Strengths: Semantic segmentation mentioned above is highly useful in cancer diagnosing through imaging. It enables
FCNs to conduct a detailed identification of cancerous regions within liver images at the pixel level. This way, FCNs
also automate the segmentation process, which saves time and reduces the potential for human error in manual
segmentation.
– Limitations: Similar to CNNs, FCNs require large annotated datasets for training to achieve high performance, which
can be a limiting factor in liver cancer diagnosis. Computational complexity is also high for FCNs, questioning their
clinical applicability.

3.2.4 Support vector machine (SVM)

SVMs are a classification tool and a supervised learning ML algorithm used in classification tasks. Similar to other algo-
rithms, SVMs are trained by exemplar data, and can then use this knowledge to make new decisions. Specifically, SVMs
are trained to assign labels to objects. Because SVMs inherently work with linear functions, they work best and most
commonly for binary decision-making and classification; however, modifications can be made to change this. SVMs are
put at the end of AI models, where segmented images and highlighted features are input into the SVM for it to classify
the object–in our case, the type of liver cancer [48]. Figure 5 shows the output of a SVM model for a binary classification
problem. This is further explained in the below paragraphs.
The blue and orange data points signify opposite classes. Maximum Margin Hyperplane: The optimal hyperplane that
is equidistant between the positive and negative hyperplane. The goal of this classifier is to maximize margins, therefore
making clear distinctions between the two classes. Positive Hyperplane: The boundary before the positive class data
points. Negative Hyperplane: The boundary before the negative class data points. Support Vectors: These are data points
closest to the optimal hyperplane (shown in white) and decide the position and orientation of the hyperplane.
In our context of liver cancer diagnosis, we have presented briefly the strengths and limitations of SVMs:

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

Fig. 5  Structure of an SVM

algorithm

– Strengths: SVMs are effective even when working with smaller datasets, as they can still perform well in binary clas-
sification tasks related to cancer diagnosis. This makes them a good choice for most cancer diagnostic models. SVMs
can also effectively handle high-dimensional data, making them useful for complex feature sets taken from imaging
and patient data.
– Limitations: Although SVMs are extremely effective in binary classification, liver tumors have variants including differ-
ent cancer types and benign tumors. This may limit their applicability in non-binary multi-class cancer diagnosis with-
out significant modifications. Also, SVMs often require manual feature selection, making them more labor-intensive
and creating the risk of not capturing all relevant features when diagnosing.

3.3 Transfer learning models

Transformers are a form of deep learning models that have started to rise in use of medical diagnosis. Transformers are
popular as they utilize self-attention mechanisms to enable themselves to weigh the relative importance of different
parts of the training and input data. This means that they are able to interpret what data and features are more significant
for making a final decision, such as a diagnosis. Self-attention mechanisms enable transformers to capture long-range
dependencies–relationships between elements of the data that are spread apart–unlike traditional neural networks.
Similarly, transformers excel in utilizing global contextual information within datasets. To exemplify this in our context,
in a medical image, accurately segmenting an organ requires understanding the overall structure and relation of the
organ to surrounding tissues and other organs. This is a benefit of using transformers [49, 50]. Below, we have elaborated
on commonly used transformers in the diagnosis of cancers using AI.

3.3.1 Medical transformer (MedT)

MedT is a specialized Transformer type designed for medical image segmentation. As mentioned above, as it is a trans-
former, MedT leverages self-attention mechanisms, allowing it to capture long-range dependencies and global contex-
tual information within images and other data. In our context of liver cancer diagnosis, we have presented briefly the
strengths and limitations of MedT below [51]:

– Strengths: It can capture long-range dependencies and thus can effectively model global contextual information to
capture anatomical structures and subtle features essential for accurate cancer diagnosis and differentiation (crucial
for liver cancers). Transformer architectures, including MedT, are highly scalable. They can handle large datasets effi-
ciently, making them suitable for large medical screening tasks.
– Limitations: MedT requires significant computational resources due to the complexity of Transformer models. This
can limit its application in real-time clinical settings. Additionally, the training of Transformer-based models typically
requires large, annotated datasets. In medical data, obtaining such datasets can be challenging due to the lack of
annotated medical images.

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

3.3.2 U‑Net variants

U-Net is a type of CNN architecture designed especially for the segmentation of medical images. Over time, as they have
gained popularity, various U-Net variants have been developed to enhance performance for specific tasks. They are
mainly known for the simplicity and efficiency they add to the process of medical imaging. U-net is also special in terms
of handling similarities between medical images. The model employs an encoder-decoder structure with skip connec-
tions that facilitate the transfer of detailed information from the encoder to the decoder, thus improving segmentation
accuracy and its ability to identify unique images [52]. In our context of liver cancer diagnosis, we have presented briefly
the strengths and limitations of U-Net below.

– Strengths: U-Net and its variants are generally simpler and less resource-intensive than Transformer-based models.
This makes them easier to train and deploy even with fewer computation resources. Also, it performs well even with
limited training data, which is common in medical applications. U-Nets have a proven track record of effectiveness in
a wide range of medical imaging tasks, including liver cancer detection. Their consistent performance makes them a
reliable choice for many applications.
– Limitations: U-Nets primarily focus on local features, which can be a limitation when capturing global context and
long-range dependencies. This can affect their ability to accurately segment complex structures compared to Trans-
former-based models. They also sometimes struggle with larger datasets and high-resolution images.

3.3.3 TransUNet

TransUNet combines the strengths of U-Net, a known CNN-based architecture, with the capabilities of Transformers. This
hybrid approach integrates a Transformer into the U-Net structure, improving the model’s ability to represent features
by efficiently using both local and global contexts. This ability significantly improves model precision and differentia-
tion capabilities as it aids in the identification of tumor regions in images [53]. For example, a study using TransUNet for
liver tumor segmentation reported a significant improvement in segmentation accuracy compared to traditional U-Net
models [54]. It also aids in delineation of tumor, reducing manual load for users.

– Strengths: The hybrid structure of TransUNet gives the model the strong local feature extraction of CNNs with the
global context understanding of Transformers. This helps have good precision while also a spatial understanding
of the cancerous region, in our case the liver parenchyma. TransUNet can be tailored to different medical imaging
modalities, making it more applicable in clinical scenarios irrespective of availability of imaging tools.
– Weaknesses: The integration of Transformers into U-Net increases the model’s complexity, which can complicate
training and deployment. This can make implementation difficult and expensive.

3.4 Fuzzy logic for AI‑based diagnostic systems

Fuzzy logic is another computational technique used when creating AI-based diagnostic models. The purpose behind
Fuzzy logic is to better replicate the uncertainty behind cases in the real world. Unlike Binary or Boolean computing, in
which 0 s and 1 s are used to encode all outcomes, Fuzzy logic allows us to create different “degrees of truth”.
When dealing with complex data, Fuzzy logic is commonly used to replicate human thinking and cognition. It can be
extremely helpful in handling machine outcomes based on numerous input parameters. In the context of deep learn-
ing, Fuzzy logic can enhance model decision making by allowing them to handle imprecise inputs and make decisions
based on degrees of truth rather than binary decisions. More specifically, liver cancer detection using deep learning often
involved analyzing complex imaging data, where variations in tumor appearance and boundaries between organs can
make it difficult to conduct binary classification on the type of cancer. Integrating Fuzzy logic with deep learning can
allow for more nuanced decision-making, providing probabilities rather than absolutes–leading to improved sensitivity
and specificity for liver cancer detection. For instance, Fuzzy logic can help identify the precise stage of a cancer, or the
likelihood for a particular case [55–57].

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

4 Literature survey

The below subsection first shows a comparative analysis between popular imaging techniques used for training AI
models. Furthermore, we provide reasoning behind the potential of 3D imaging datasets. This review also consists of
original studies analyzed, which were chosen based on the inclusion criteria shown above. 8 of the above papers con-
tain some variants of CT datasets, as it is the most popular choice in AI models. 2 papers contain MRI datasets, and then
miRNA followed by DNA methylation dataset based studies have been analyzed. The difference in dataset types depicts
the ratio observed during research.

4.1 Imaging for deep learning AI models

In the past, the most efficient non-invasive method for liver tumor diagnosis has been some form of imaging (one of the
3 previously explained) [58]. Imaging is used because it can give information about the precise location of the tumor/
mass and its size and shape. This is also extremely important in being able to differentiate between types of liver cancers.
It’s shown that there are particular important features in imaging that are more likely to appear in HCC than in ICC, for
example. These include intralesional fat, lesion washout and the presence of a capsule [59].
This shows that accurate and efficient imaging is extremely useful for diagnosis of liver tumors. Differentiating between
types (HCC vs ICC) also gives information for the prognosis of the tumor.
When training Deep Learning AI to diagnose and classify liver masses, CT and MRI are the preferred choices for input
training datasets. Ultrasonography (US) is fast, but has other disadvantages that make its use limited: US is unsatisfactory
because of its low sensitivity to small liver masses and the dependency on a manual US operator [60]. CT and MRI are
both automated and done through machines. The two also provide a higher resolution and level of detail—enabling
higher precision in distinguishing normal and abnormal tissue/patients [61]. According to research by Zhen et al. [60],
enhanced MRI is the best tool for diagnosing liver masses because it has various mechanisms for tissue contrast and has
the ability to clearly check auxiliary imaging features. MRI also allows good categorization of small tumors (≤ 3 cm) [61].
Even with this, MRI is not used widely unlike CT and US because of the high cost and time associated with it.
Additionally, MRI and CT can be combined with contrast agents to assess blood flow in the liver, which is an important
factor to help distinguish benign and cancerous lesions. Digging deeper, it was observed that between these two (MRI
and CT), CT was the most opted-for option. Alongside the high resolution, another advantage of the use of CT is that
it allows for similar or better evaluation of extrahepatic tissue, including bones, bowel, lymph nodes, etc. as MRI while
being quicker and more affordable [62]. It is therefore recognized as the preferred medium for imaging cancers in the
liver despite the ionizing radiation exposure to the patient [61]. Because of this, most of the studies analyzed use a CT-
orientated dataset.

4.2 2D and 3D datasets for deep learning AI

Both CT and MRI scans are inherently 3D, but CT is often represented as a continuous series of 2D cross-sectional slices
of the 3D scan. This is done to make it easy for radiologists to interpret the data. Also, 2D data requires significantly less
computational complexity, power, and storage, thus is the preferred choice for the localized analysis of a disease region
where a single cross-sectional view is sufficient. However, 2D images create a lot of restrictions. They reduce our ability to
understand spatial analysis, where experts check for the spatial relationships between structures in the body. 3D images
give a more holistic view and help to view anatomical structures more clearly. In 2D images, overlapping structures are
harder to differentiate and identify as well.
With time, a new trend has come where 3D data is being used more often for training and testing AI models for dis-
eases. For CT scans, a 3D dataset is created by stacking 2D slices, resulting in a volumetric representation of the imaged
area. In the case of cancers, a volumetric representation and a 3D view make the shape and location of the tumor more
easy to locate. 3D data also helps monitor tumor growth and vascularization. For example, they often incorporate blood
vessel distribution around the tumor, which helps in differentiating cancers.
Moreover, 3D datasets are definitely more comprehensive in comparison to 2D; however, algorithms incorporating
3D data tend to be significantly more complex, increasing training and testing time and computational complexity. This
may reduce a model’s ability to be used in day-to-day clinical applications.

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

4.3 Fuzzy logic‑based diagnosis for liver cancers

Fuzzy logic has a lot of potential in the field of liver disease diagnosis. Ahmad et al. [63] showed that fuzzy systems can
be used to accurately classify liver diseases. Ahmad et al. [63] proposed a multilayered Mamdani fuzzy inference system
for diagnosing a liver virus (Hep-B) using a fuzzy-based inferential system. Similar to the potential fuzzy logic applica-
tions listed earlier, this model is able to classify different phases hepatitis B because of fuzzy logic. The system showed
a 92.2% reliability. Furthermore, [64] uses a Mamdani fuzzy inference method to create a 95% accuracy model for liver
fibrosis detection. This study explained that fuzzy logic can be used to deal with inaccuracies and vagueness in patient
historical data. In addition, by using fuzzy data it was possible to classify patients’ health into different classes. Mirmozaf-
fari [65] utilizes an expert system for diagnosing liver disease. These systems are similar in function to Fuzzy systems as
they store experts knowledge and use it to process information and make decisions (generate outputs). Lastly, another
study, [66], made use of a Fuzzy Cluster Means (FCM) system to diagnose liver diseases. Using FCM aided in evaluation
and classification of patients into multiple classes of liver diseases.
The integration of fuzzy logic into liver disease diagnosis demonstrates significant potential in increasing accuracy
and reliability. Various studies above demonstrate its effeciveness. Fuzzy logic allows for impressive classification as it
becomes easier to identify stage of disease and differentiate between classes of diseases/cancers. Fuzzy-based models
certainly have potential in liver cancer diagnosis; however more research in the field is needed.

4.4 Research analysis

This subsection highlights key research studies on liver cancer detection and classification using a variety of dataset
types and deep-learning AI models. Table 2 provides a description of different studies along with their sample size, case
and objective.
Table 3 presents the research analysis done on recent studies. Here, we have analyzed the challenges associated with
the various Deep Learning algorithms used in the proposed models, while also noting their future scope in medical
practices.

5 Benchmark AI‑based framework

The ability of artificial intelligence ML and DL to make complex decisions based on empirical data presented to them
has shown to be significantly helpful in modeling and aiding in complex problems like liver cancer and other diseases. AI
tools are constructed following specific standard processes to work efficiently for any application. These tools then aid in
faster, more efficient and reliable diagnosis and prediction of diseases. A benchmark construction procedure of AI-based
diagnostic systems is shown in Fig. 6. This benchmark framework aims to address commonly occurring limitations in
previous works, such as balancing, overfitting, etc. This structure is not a complete model, rather, can serve as a refer-
ence point for future models and researchers can build off of this. This is important as it addresses common limitations.
Exploratory data analysis (EDA) refers to the data exploration process. It is used to better understand the information
presented in datasets by highlighting important characteristics and is used to spot patterns and relationships between
variables. Techniques such as correlation analysis, filtering out non-essential data, and identifying outliers are commonly
used to extract meaningful insights from the data. Next, we go on to preprocessing techniques.
First, Data Curation. In this step, data is managed and organized by the curators to make it easier for users to analyze
the data. Further investigation of the data is done through analysis of associated repositories. Class balancing is then
done to the dataset when the instances of each class are not equally distributed within the dataset (i.e. class imbalanced
data). The imbalance Ratio (IR) is computed using the size of the minority class (­ Mi) and the size of the majority class
­(Ma). IR is calculated using the Eq. (1).
Mi
IR = (1)
Ma

This is extremely important to acknowledge because training AI on such data will make the model biased towards the
majority class, and show poor performance for the minority classes. Oversampling (increasing the number of instances in
the minority class synthetically or by duplicating) and undersampling (decreasing the number of instances in the majority

Vol.:(0123456789)
 Table 2  Research analysis
Authors Dataset type Sample size Cases Objective
Review

Zhen et al. [60] Contrast-enhanced MRI, unenhanced 31,608 images 1210 Dynamic contrast-enhanced MRI allows
MRI for the most holistic and compre-

Vol:.(1234567890)
hensive observation for differential
diagnosis of liver tumors. However, the
subjectivity of doctors poses a prob-
lem. Deep learning is used to combat
this. Assisted diagnosis of liver tumors
by categorizing them into 7 categories
using enhanced, and unenhanced
images and clinical data (like text)
Gao et al. [71] Multi-Phase Contrast-enhanced CT Not mentioned 499 To create an AI diagnostic assistance
system for doctors to help differentiate
malignant hepatic tumors based on
multi-phase CECT images
Wang et al. [72] Dynamic CT Not mentioned 234 (157/77 for train/test) To integrate the siamese mechanism
into (cross-contrast NNs) CCNNs, allow-
Discover Applied Sciences

ing them to predict like traditional

CNNs, while taking advantage of the
siamese twin network mechanism for
better pattern recognition for diagno-
sis of HCC and ICC
Yang et al. [73] Contrast-enhanced CT Not mentioned Training/testing—198/85 To aid in prognostication of HCC recur-
rence, recurrence-free survival (RFS)
(2024) 6:508

and overall survival (OS) of early-stage

HCC patients through the use of
radiomics
Yasaka et al. [67] Dynamic CECT 1068 images pre-agumentation. 55,536 460 patients (for training). 100 patients To develop a CNN to analyze dynamic
images post augmentation (testing) contrast-enhanced CT scans and
automatically classify liver masses into
distinct categories (HCC, Malignant
tumors other than HCC, indeterminate
masses, hemangiomas, cysts). The
study aimed to overcome previous
limitations in the field in regard to
differentiating indeterminate lesions.
One example of this is the similarity in
attenuation of early HCC and dysplas-
tic nodules leading to false diagnoses)
| https://doi.org/10.1007/s42452-024-06218-0
 Table 2  (continued)
Authors Dataset type Sample size Cases Objective

Das et al. [74] CT 225 75 patients (46 male, 29 female) The study aimed to develop a new
automated method for liver cancer
detection in CT images. Thus, an auto-
mated CAD system was developed for
decision-making processes for cancer
diagnosis. New technique introduced:
Watershed Gaussian-based deep
Discover Applied Sciences

learning
(WGDL) technique. For cleaner images
and improved segmentation and clas-
sification
Rajkumar et al. [75] CT 5 Not mentioned The primary aim of the study was to
develop a more accurate method for
automatic liver tumor detection in CT
(2024) 6:508

images by combining two popular

methods, KBC methods for segmenta-
tion and SVM classifer
Chlebus et al. [76] CECT 131 CECTs Not mentioned This study aimed to create a U-net
based FCN model with object-based
postprocessing for successful and
automated liver tumor segmentation
in contrast to enhanced CT scans. The
model aims to utilize full 3D volu-
metric information for more efficient
classification of tumors
Ben-Cohen et al. [77] CT NA 20 patients with 68 lesions (training). The research aimed to develop a novel
14 patients with 55 lesions (testing) model for the detection of liver lesions,
specifically for increasing success-
ful diagnosis and detection of small
lesions and metastases in CT scans
Trivizakis et al. [68] Diffusion-weighted MRI 130 DW-MRI scans Not mentioned This study proposes a novel and rather
unpopular idea of utilizing 3D CNNs
as feature extractors and SVMs as
classifiers for diagnosis of liver cancers
in a diffusion-weighted MRI dataset.
| https://doi.org/10.1007/s42452-024-06218-0

The study aimed to combat difficulties

in previous studies in distinguishing
between primary and metastasized
tumors in the liver when using 2d
CNNs

Vol.:(0123456789)
Review
 Table 2  (continued)
Authors Dataset type Sample size Cases Objective
Review

Sayed et al. [69] Micro RNA (miRNA) Not mentioned Not mentioned The study’s primary objective was to
identify 3 novel miRNA biomarkers

Vol:.(1234567890)
that could be used as early diagnostic
markers for HCC. The study utilizes
these biomarkers and their expression
profiles to propose a highly accurate
machine learning model that predicts
HCC based on miRNA profiles of
patients. The model is robust as it also
incorporates hematological, biochemi-
cal, and microbiological markers
Bai et al. [70] DNA Methylation sites 96 + 402 + 108 liver samples Not mentioned The study aims to identify DNA methyla-
tion markers for 2 purposes: differ-
entiating primary liver cancer (PLC)
with normal tissue, and differentiat-
ing between HCC and ICC. The study
Discover Applied Sciences

develops machine learning models

that differentiate between the 4
categories above with high accura-
cies based on methylation data of the
patient
Liu et al. [78] mRNA, lncRNA, miRNA, CNV, and SNP 320 + 320 + 321 + 324 + 313. 302 sam- Not mentioned The study aims to develop a more accu-
ples for training rate prognostic model for HCC using
(2024) 6:508

multi-omics data. This is done by inte-

grating various omics data, including
mRNA, lncRNA, miRNA, CNV, and SNP
information. These datasets were used
since they clearly represent biological
phenomena behind HCC
Lu et al. [79] Gene expression profiles 225 + 268 + 166 + 48 + 91 = 798 Not mentioned The goal of this study was to identify key
pathways and genes involved in the
pathogenesis of HCC through cross-
species comparison and bioinformat-
ics analyses
| https://doi.org/10.1007/s42452-024-06218-0
 Table 2  (continued)
Authors Dataset type Sample size Cases Objective

Schmauch et al. [80] Ultrasound 367 images Not mentioned The study developed a deep learning
model for the detection and clas-
sification of focal liver lesions into
the following categories: malignancy,
angioma, hepatocellular carcinoma,
focal nodular hyperplasia, cyst, or
metastasis–done using ultrasound
Discover Applied Sciences

imaging data. It was designed to

perform two main tasks: detecting
abnormalities and classifying lesions
as benign or malignant
Suyama et al. [81] FDG-PET vs. EOB-MRI scans Not mentioned 34 liver metastasis cases The study investigates the diagnostic
ability of a deep learning image recon-
struction (DLR) method in detecting
(2024) 6:508

liver metastases cases using FDG-PET

imaging. The primary purpose of the
study was to compare the efficacy of
the DLR in liver metastases diagnosis
against the established EOB-MRI find-
ings
Kanan et al. [82] CT Not mentioned 121 patients with liver metastases The primary objective of the study was
to determine whether Deep Learning
Image Reconstruction (DLIR) can aid in
the detection of liver metastasis cases
in the CT scans of patients. The detec-
tion was performed by radiologists
who utilized both DLIR and a more
traditional adaptive statistical iterative
reconstruction (ASiR-V) method. The
conspicuity of the 2 were compared to
determine which method was better
Shen et al. [83] Gene expression profiles NA 362 patients with HCC The primary goal was to identify a
unique gene signature that could pre-
dict the recurrence of HCC in past-HCC
patients. The study aimed to create a
| https://doi.org/10.1007/s42452-024-06218-0

cost effective gene panel for clinical

use since whole-exome and whole-
genome sequencing is expensive. The
ML model was made after analysing
gene mutations from databases and
comparing that with whole-exome
sequencing of 10 patients

Vol.:(0123456789)
Review
 Table 3  Description of challenges encountered and the possible future outcomes for studies analyzed
Authors ML/DL model used Challenges Future scope
Review

Zhen et al. [60] CNNs High number of false negatives for HCC diagnosis. Adding a variety of unique cases in the training set
Radiologists performed visibly better for the cat- will help reduce false negatives. Better utilization of

Vol:.(1234567890)
egory metastatic malignant tumor clinical information will help for improved meta-
static malignancy categorization. “accurate and
time-saving assisted-diagnostic strategy for liver
tumors”. The use of unenhanced images in training
allows for use in the absence of contrast agents
Gao et al. [71] STIC model with integrated deep CNN and gated Small training set and thus difficult for atypical Since the STIC-assisted doctors set showed a
RNN cases and generalizability. Increased training time significant increase in accuracy and sensitivity in
because of multi-phase CTs. The model alone had a the diagnosis of malignant hepatic tumors (26.9%
low accuracy (ranging between 70–80%) but STIC- increase in sensitivity for ICC), the model has scope
assisted doctors set showed high accuracy for application in clinical settings as a CAD tool that
is relatively easy to interpret
Wang et al. [72] Siamese cross contrast neural network (SCCNN) The performance of SCCNN is influenced by the SCCNN is a viable choice in situations where less
choice of the base model, and more complex base medical data is available. The model shows good
models may not always lead to better results. performance as it is able to extract more discrimi-
Discover Applied Sciences

SCCNN feature extraction is distributed throughout native features and work with lesser medical data.
the entire liver, showing increased reliability in This shows its ability to be a CAD tool in clinical
non-lesion areas. This hinders diagnosis of larger applications
tumors
Yang et al. [73] Radiomics. Pretrained CNNs, SVM classifier Retrospective study, therefore, training and testing Model showed satisfactory MVI (microvascular inva-
was done on past data, thus making generalizabil- sion) prediction and high efficiency overall. The
ity in present times difficult. Small dataset as well model also doesn’t require manual delineation. This
(2024) 6:508

automation and accuracy allow it to aid in clinical

decisions on patients with early HCC and to help in
the recurrence of tumors and even mortality
Yasaka et al. [67] CNNs The model showed low sensitivity for diagnosis of The model showed high performance regardless of
indeterminate lesions including early HCC and dys- what was shown around the liver masses. This might
plastic nodules likely due to the subtle difference have been achieved by training with several liver
shown in the slices and limited training data for masses in various sections of the liver and because
those masses. The sensitivity for rarer masses was of high augmentation of image sets. Additionally,
lower for the same reasons. Models like CNNs also the model showed high performance for most
experience low performance because of overfitting masses encountered in daily clinical practice, mak-
caused by large datasets ing it applicable in real-life settings
Das et al. [74] Watershed Gaussian based deep learning. DNN Use of watershed transform and Gaussian mixture The model showed a 99.38% accuracy and 100% sen-
classifier model makes the model have high computational sitivity in detecting liver cancers in the testing set.
complexity, making it difficult for regular clinical This great success makes the model viable for use in
applications. Small dataset. The model also strug- clinical diagnosis and decision-making process for
gles with calculating the volumetric size of lesions diagnosis of liver cancers. Furthermore, making 3-D
mesh structures from different slice of the images
will help calculating volumetric size of the image
| https://doi.org/10.1007/s42452-024-06218-0
 Table 3  (continued)
Authors ML/DL model used Challenges Future scope
Rajkumar et al. [75] SVM, KBC (knowledge-based constraints) Extremely small test size of only 5 images. SVM clas- The model has extremely high accuracy of above 98%,
sifier is not optimal because of the large datasets. and above 98% sensitivity and specificity. The model
The proposed model also requires excess manual should be tested with a larger set to check generaliz-
interaction and has a slow computational speed ability and be trained with more images as well
Chlebus et al. [76] 2D FCNN and Random forest classifier (RFC) The use of voxels to utilize volumetric information The model gave results comparable to a human
led to the high false positive rate in the study. The expert (77% detection), however, is capable of
Discover Applied Sciences

model also showed a low success rate for detecting detecting bigger lesions with better accuracy than
smaller tumors, and an overall inferior detection smaller ones. More 3D volumetric information
(not segmentation) success in comparison to should be utilized for better accuracy
humans (63% vs. 92%)
Ben-Cohen et al. [77] FCN, Sparsity-based classification The FCNs performance was worsened because of Higher resolution and thinner slices would improve
slice thickness of the CTs. Along with this, the use the detection rate, especially for small lesions. The
of the automatic liver segmentation tool reduced model showed a fully automated convolutional
(2024) 6:508

the true positive rate slightly. Inaccuracy of this network with sparsity-based false positive reduction
tool caused some small proximal detected lesions for the detection of liver lesions, specifically liver
to be merged metastases in CT. The model presented a high 88.2%
true positive rate, and a reduced false positive rate
with sparsity classification. This makes it viable for
trial in clinical applications
Trivizakis et al. [68] 3D CNNs, SVM classifier In DW-MRI, primary and metastasized tumors exhibit When compared to 2D CNNs, which were similarly
slower signal decay at high b-values, giving this trained, the model displayed a higher accuracy for
study a low accuracy when distinguishing between all tests (85.5%/71.9% for primary vs. metastasized
the 2. Small dataset. 3D CNNs increased compu- tumors for 3DCNN/2DCNN). The model also removes
tational complexity making clinical applications the need for processes like delineation, cropping,
difficult and ROI annotations reducing time and making it
viable for clinical application if computational com-
plexity is reduced
Sayed et al. [69] BAVO (binary African vulture optimization algo- Firstly, the model was trained and tested on a small The model should be tested and trained on a larger
rithm) and SVM dataset, as acknowledged by the authors, and this cohort to make it applicable in clinical use. However,
may affect the generalizability. Also, the dataset the accuracies for all 3 miRNA studies were high in
only uses Egyptian patients’ profiles, reducing comparison to similar studies. This study therefore
diversity in the training data. Next, out of the 3 encourages research on the miRNAs found to be
miRNAs tested, 2nd and 3rd miRNA showed low linked to early-HCC. Lastly, the novel BAVO algo-
accuracy and performance due to the limited avail- rithm introduced in this paper for an efficient feature
able information on them. This shows a limitation selection and optimal parameter selection has lots
| https://doi.org/10.1007/s42452-024-06218-0

of miRNA datasets since they are still undergoing of potential implications in the field as it resulted in
research and information is not readily available for high accuracies in this paper
many miRNAs

Vol.:(0123456789)
Review
 Table 3  (continued)
Authors ML/DL model used Challenges Future scope
Review

Bai et al. [70] Random forest The most significant challenge in the study was its The problem of differentiating between cancer
dependency on biopsies, which are traumatic to types in the liver is a big one (because its prone to

Vol:.(1234567890)
the patients. This means that for clinical applica- metastases), but the use of methylation biomarkers
tions of this model, taking relevant tissue samples showed great accuracy for classifying HCC, ICC and
from patients will be traumatic and thus not the PLC in the liver. This could also be used to differenti-
best approach. The invasive approach utilized is ate metastatic tumors, cysts and lesions from PLCs
the challenge, and so the efficacy of the methyla- as well, resolving a common problem with diag-
tion biomarkers may not be the same in actual nosing liver cancers. Since the authors plan to test
patients the model again on peripheral blood samples, the
trauma associated with this method of diagnosis
may be resolved with future research
Liu et al. [78] Multi-omics There was no external validation for 3 of the single- The multi-omics model exhibited better performance
omics models due to the lack of independent (Concordance index based) than all single-omics
datasets. Similar problem persists for the multi- models. Multi-omics data has great potential in
omics model as well. Next, the final multi-omics terms of accuracy and prognostication of liver
model includes 30 + biomarkers and thus may be cancers. External validation is necessary for the
Discover Applied Sciences

difficult to use in clinical setting. The success of the model to go for clinical use. As mentioned earlier,
model strongly depends on patient willingness to simplification of the model is necessary to enhance
undergo sequencing practicality for patient comfort and easier usage
Lu et al. [79] Gene set enrichment analysis (GSEA) and meta- The study acknowledged that there was a small The study identified the cell cycle pathway as signifi-
analysis sample size, which could lead to unreliable vari- cantly enriched across multiple datasets. Specifically,
ance estimation and thus a higher false positive cdc25a was noted as a key gene in HCC tissues com-
rate. This may also cause accidental overlooking pared to normal tissues across the 3 species. Further
(2024) 6:508

for some significant differences in gene expression validation is required for the identified pathways;
levels of different samples. The use of different however, these have diagnostic potential for HCC.
samples and experimental platforms may cause Integration with a machine learning model may
the datasets to be unstandardized, potentially allow for genome-wide data interpretation leading
reducing the robustness and generalizability of the to more identified pathways linked to HCC
data
Schmauch et al. [80] ResNet50, supervised attention mechanism There was a clear imbalance in the distribution of The model uses a novel approach to limit the annota-
the lesion types within the dataset, with some (ex: tion time for radiologists. Before further implemen-
HCC) being underrepresented. This limited expo- tation, traditional more precise annotation methods
sure to diverse cases reduces the generalizability of should be tested to see if they result in stronger
the model. Additionally, the authors acknowledge accuracies. Simultaneously, the new annotation
that some annotations may be unrepresentative of approach here also may have implications for other
the true nature of lesion in the ultrasound image: imaging-based deep learning models. Moreover, a
labels such as “HCC” and “Metastasis” don’t make larger external validation set should also be consid-
sense to provide as they are difficult to diagnose ered, which was also acknowledged by the authors
with ultrasound alone. A larger testing dataset is of the study
also required to validate the model
| https://doi.org/10.1007/s42452-024-06218-0
 Table 3  (continued)
Authors ML/DL model used Challenges Future scope

Suyama et al. [81] Deep learning image reconstruction (DLIR) The study only involved 34 diagnosed cases, not Through the low p value and possibly low sensitiv-
making the results very reliable. Additionally, ity, it is unjustified to say that DLR is significantly a
although the DLR method certainly showed a better approach to diagnose liver cancer metasta-
higher diagnostic score, the p value was only 0.14, ses; however, the research does show potential for
indicating that it was not statistically significant. DLR methods in the future, as they did increase the
Lastly, the EOB-MRI findings showed some nodules diagnostic score in the study. Larger datasets and
that were missed by the DLR, indicating a lower improvements on the sensitivity should be points
Discover Applied Sciences

sensitivity for further research

Kanan et al. [82] Deep Learning Image Reconstruction (DLIR) This was a retrospective study that used only a single The high-strength DLIR statistically increased the
CT and only 1 manufacturer’s DLIR. This makes it detection and conspicuity of liver metastases
difficult to generalize the results. Additionally, the compared to ASiR-V. This shows its implications in
inclusion criteria for the data used added limita- improving differentiation of PLCs and SLCs. Being
tions to the study: the study only included patients used as a pre-processing tool, DLIR is easier to imple-
with liver metastases, meaning that the results only ment in clinical scenarios as it is not fully automatic;
(2024) 6:508

show DLIR to improve sensitivity, not specificity rather, it is an assistive tool for radiologists
Shen et al. [83] Decision tree versus SVM The mode’s verification on a small sample size of The 2 models tested showed accuracies of 74.19%
10 patients limits the generalizability. Larger (decision tree) and 80.65% (SVM). Although the SVM
validation set is needed, given the complexity and showed higher accuracy, the authors noted that the
heterogeneity of HCC. The study noted differences decision tree model is more interpretable and thus
in gene mutations between races (asian vs non- should be focused on to improve in future research
asian population) which could affect the model’s as understandable models have greater potential for
applicability in real-world scenarios with diverse clinical implementation. Additionally, the gene sig-
patients natures need to be validated with a larger, multi-eth-
nic cohort to confirm the performance. Multi-omics
approach can also be taken to extend this research
by combining with other data, such as miRNA
| https://doi.org/10.1007/s42452-024-06218-0

Vol.:(0123456789)
Review
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

Fig. 6  Benchmark AI-model

framework

class to match the minority class) are two major techniques used to achieve class balancing. Hybrid techniques can also
be used. Another preprocessing technique done to the dataset is removing redundant and noisy features. Because these
add unnecessary load to the training data when training an AI-based model. This reduces the model’s accuracy because
relevant features are overshadowed, and also reduces its generalizability.
Data standardization is a process used to standardize the independent features (columns) in the dataset into the [0, 1]
interval. It reduces the skewness of the data distribution by shifting the individual attributes to have an average of zero
and unit variance. The standardization ­(Zn) is attained through the z-score formula below, Eq. (2).
(n − 𝜇)
Zn = (2)
𝜎
Here n represents the data instance while μ and σ are the mean and standard deviation of the feature respectively.
Lastly, data splitting: This is a step done to measure the performance of the model developed. The dataset is split
into 2 parts, where one is used to train the model and the other is used to test the model. This step evaluates the
performance of the model on unseen, new data.

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

Then we can go ahead to add the classifier. This step is highly variable depending on the type of data used. Dif-
ferent models are chosen for their performance in different scenarios. For example, neural networks require large
datasets and training. The choice of the classifier is critical as it will determine the performance. Then we tune the
classifier–selecting proper hyperparameters in order to maximize the classifier’s performance. Optimizing hyperpa-
rameters is also done to improve the performance of the model overall. For example, Grid search to find optimal
hyperparameters combination for the best diagnosis and classification. The performance of the model is then meas-
ured and evaluated on various parameters, such as accuracy, sensitivity, etc. A more elaborate explanation of these
parameters can be seen in Sect. 6.1. Although some parameters like accuracy are usually applicable, the choice of
parameters chosen also depends on the class balance within the dataset.
Model deployment refers to the process of integrating a trained DL/ML prediction model into a productive environ-
ment as a decision-making diagnostic assistant. The deployed model is then validated, in terms of performance, through
external means (such as clinics and other medical facilities). Once sufficient validation is done and the model is proven
to be useful and accurate, it can then be utilized in general clinical practice. Once validated, models are turned into ML
APIs to be available for download and use by the public.
Although our paradigm for future models offers a path forward, it is critical to recognize existing constraints. These
include issues with data quality, interpretability of the models, and possible biases in training datasets. These problems
must be resolved to successfully integrate models into clinical practice. We can get beyond these obstacles and raise the
acceptance of AI technologies in healthcare settings by increasing the diversity of data and creating more transparent
models.

6 Discussion

This study effectively analyzed the performance, features and techniques used in Deep Learning computational models
for the prediction of liver cancers. Deep Learning has grown to be a tremendous help in medicine, especially in the field
of disease diagnosis. These models have been shown to play a great role in helping doctors’ decision-making in medi-
cal processes. The highest sensitivity achieved across the models was 100%, and 99.38% highest accuracy. However,
another study showed 97% accuracy, and more studies showed accuracy and precision to be greater than 90%. Various
deep learning techniques have been utilized, including CNNs, SVMs, FCNs, and other novel combinations such as the
SCCNN discussed. Deep Learning models are beneficial because they meticulously search the input data for patterns,
extracting features indicative of cancers within the liver. This ability of DL AI enables recognition of tumors in the liver
with ease and minimal manual interaction.
Furthermore, AI has the potential to significantly decrease diagnosis time, allowing for a proactive approach to treat-
ment methods. AI also reduces the involvement of potential viewer biases when diagnosing liver cancers non-invasively,
such as through imaging [84]. This is possible because they rely purely on previously annotated and factual data and are
standardized to allow for reliable diagnosis.
Before we move on, its important to acknowledge the logistical and other difficulties associated with the clinical
implementation of AI-based predictive models. The high cost of AI technologies can be a significant obstacle. Develop-
ing, deploying, and maintaining such systems usually requires substantial financial investment. Integration with existing
medical infrastructure is another issue likely to come across since it is tedious to incorporate these models to ensure
compatibility with current electronic health records and other clinical systems. Additionally, user acceptance is crucial:
healthcare professionals may be reluctant to adopt new technologies because of various concerns about the reliability
of AI-driven models and the need for extensive training to effectively utilize these tools.

Table 4  Confusion matrix Actual condition Positive condition Negative

condi-
tion

Positive condition TPn FPn

Negative condition FNn TNn

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

6.1 Research investigation 1: evaluation parameters for AI ML/DL predictive models

Evaluative parameters are the measurements used to judge the diagnostic and predictive ability of AI models. These
metrics are also crucial to compare results between models and even compare to human experts to understand the
real-world practicality of the model at hand. They are also used to understand the advantages and disadvantages of
the models. The evaluation parameters most commonly used and those that are mentioned in this paper are described
below. It is significant to analyze the confusion matrix before understanding the evaluation parameters. The confusion
matrix, as shown in Table 4, plays a significant role in statistical classification. The layout of the table allows to visualize
the performance of a model.

• Area under the curve (AUC): AUC refers to a measure of the ability of a model to distinguish between given classes,
and is a range from 0 to 1. 0 represents exceedingly bad, whereas 1 represents very good performance. AUC curve
is a measure and summary of the ROC (receiver operating characteristics) curve, which shows the true positive rate
against the false positive rate for readers to compare.
• Accuracy ­(Accn): Accuracy can be characterized as the ratio of correctly identified cases by the model to total cases,
as shown in Eq. (3).
TPn + TNn
Accn = (3)
TPn + TNn + FPn + FNn

• Sensitivity/Recall ­(Rn): Sensitivity is characterized as the ability of the model to correctly identify positive cases. Equa-
tion (4) shows this:
TPn
Rn = (4)
TPn + FNn

• Precision ­(Pn): Precision is the ratio of correctly identified positive outcomes to total positive outcomes returned by
the classifier. Mathematically, it is shown in Eq. (5).
TPn
Pn = (5)
TPn + FPn

• Specificity ­(Spn): Specificity refers to the ability of the model to correctly identify negative cases. Equation (6) shows
this.
TNn
Spn = (6)
TNn + FPn

6.2 Research investigation 2: common available datasets for liver cancers

In Table 5 we have presented various publicly available datasets for liver cancers that include most of the diagnostic
procedures previously discussed and additional types.

6.3 Research investigation 3: research gaps identified in current research

This section will be moving on from the challenges already discussed in Table 3. As previously discussed, the liver is a site
prone to metastases in cancer patients. Because of this, the major problem that diagnostic systems aim to overcome is
to be able to differentiate between PLCs and secondary malignant tumors in the liver. This problem not only troubles
experienced professionals, but is a major research gap in the field of Deep Learning diagnostic models; specifically, mod-
els that utilize imaging datasets (MRI, CT, US, etc.) struggle with this. This was first observed in [67]: this study mentioned
that one of the major limitations in the field is the differentiation of liver masses. In the 2D CT slices used for the study,
early HCC tumors appeared similar in appearance and attenuation with dysplastic nodules in the liver. Furthermore,
MRI screening also has this limitation. This study [68] used a diffusion-weighted MRI dataset and observed that DW-MRI

Vol:.(1234567890)
 Table 5  Descriptions of commonly used datasets for training AI for liver cancer diagnosis
Dataset name Dataset link Type of dataset Description

3D liver and liver tumor segmentation https://​www.​kaggle.​com/​datas​ets/​gaura​vdutt​ 3D CT and MRI To ensure generalizability, the dataset includes a
akiit/​3d-​liver-​and-​liver-​tumor-​segme​ntati​on/​data diverse set of patients covering a range of ages,
genders, and ethnicities. Each medical image has also
been annotated by professionals. It includes a variety
of pathological variations in tumors (size, benign,
malignant, etc.). 3D data helps train for volumetric
Discover Applied Sciences

information and real-world data

Liver tumor segmentation https://​www.​kaggle.​com/​datas​ets/​andre​wmvd/​ CT The dataset consists of 130 CT scans of the liver and of
liver-​tumor-​segme​ntati​on/​data liver lesions intended for liver tumor segmentation.
The dataset was extracted from the LiTS challenge.
The data was provided by multiple clinical sites over
the world to ensure diversity in the data
The cancer genome atlas liver hepatocellular carci- https://​www.​cance​rimag​ingar​chive.​net/​colle​ction/​ MRI, CT and PT The dataset is a free, public dataset consisting of data
(2024) 6:508

noma collection tcga-​lihc/ from 97 subjects, giving a total of 125,397 images.

Vast quantity helps the adaptability of the model
however this dataset focuses only on HCC cases so is
not feasible when trying to diagnose cancers other
than HCC
High-sensitive circulating microRNA panel for early https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​ MiRNA Profile In such datasets, profiling of non-coding RNA is done
detection of hepatocellular carcinoma acc=​GSE11​3740 using an assay for several patients. Here, serum micro-
RNA profiles of 1817 samples were analyzed, with
patients of HCC, non-cancer, hepatitis, and liver cir-
rhosis. The expression of serum MiRNA combinations
can be used as a classifier for liver cancer in patients.
Having profiles of a diverse set of diseases helps
specify which upregulations are linked to liver cancer
Gene expression profiles of human hepatocellular https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​ Gene expression This dataset consists of 31 gene expression profiles of
carcinoma acc=​GSE45​435 profiles HCC samples. The main aim was to understand the
differences in gene expression of young and old HCC
patients. Understanding gene expression in different
diseases can help link certain genes to look out for
during diagnosis
Single cell transcriptome of human intrahepatic https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​ Transcriptome Unlike most datasets, this dataset focuses on ICC. The
cholangiocarcinoma acc=​GSE18​1878 sequencing set involves single-cell transcriptome sequencing
| https://doi.org/10.1007/s42452-024-06218-0

data of 3 samples (plasma, tumor, and paracancerous

tissue) from an ICC patient. At a single-cell level, this
dataset helps understand and detect unique cell het-
erogeneity within ICC cells, and to identify variations
in cell types in different tumor environments. This
dataset can be used to understand the tumor micro-
environment in ICC patients to classify the particular
cancer. However, only 1 patient’s data is used, so its
not sufficient and generalizable

Vol.:(0123456789)
Review
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

shows a slower signal decay at high b-values for both primary and metastasized tumors, reducing the study’s precision
in distinguishing between the two. Because of the large variety of possible tumors in the liver, and the limitations that
come when using imaging as the sole diagnostic tool, misdiagnosis and false positives often come about.
This limitation also reduces the independence of the models. A recurring trend was that the accuracy of the DL
model-assisted doctors set was the highest, likely because the doctors are able to better utilize clinical information in
their decision-making, adding to the ability of the model alone.
This problem continues between benign and malignant masses as well. A high false positive rate was common
between models. Conditions such as hepatic hemangiomas, cysts, and focal nodular hyperplasia can mimic liver tumors
when using imaging techniques, giving false positives. However, it is likely that the use of 3D imaging has the potential
to reduce this because of the additional information it provides.
Another area of research that should be further explored is the use of 3D imaging data, including 3D CT and MRI, in
combination with 3D neural networks. A vast majority of the papers found and discussed utilized imaging datasets, spe-
cifically CT, but only few utilized 3D tomographic data. Reliance on 2D data presents the models with various limitations.
Firstly, 2D CNNs analyze slices of data individually, missing out on crucial volumetric information. They don’t provide
enough information about tumor morphology, including its interactions with surrounding organs and tissue, its shape,
and location. Additionally, specifically in the liver, attempts to delineate tumors spread across multiple 2D slices often
result in errors like missing out on parts of the tumor due to “multi-focal” tumor appearance [68].
Moving on, many machine learning models relied on static images as their primary datasets. This refers to simple CT,
CECT, MRI, etc. The advantages of a temporal analysis through sequential imaging techniques appeared to be underu-
tilized. Few studies implemented dynamic CT/MRI, or PET scans for training their models, even though they provide
additional information. This is likely because of the increased training time and computational complexity that comes
along with dynamic imaging; however, such techniques allow observations on the tumor behavior (see previous section
on dynamic imaging), aiding in early diagnosis and prognosis of cancers. Utilizing this data would likely provide models
with more discernible features and characteristics associated with specific cancers based on their interactions with sur-
rounding tissue–which can be observed through dynamic imaging.
The datasets used also bring about other problems. Commonly, models are trained on data from one specific medi-
cal center, and since imaging protocols like contrast phases, resolution, etc. are not always standardized across centers,
models end up doing worse in the validation set, or in real-world applications. This is a big reason behind the delay in
FDA approval for the clinical use of many AI models. It is vital that in training, models are trained on a variety of datasets
from different centers, that also take into account a diverse group of patients and diseases, in this case: benign masses,
HCC, ICC, secondary liver tumors, etc. Overfitting is also prevalent and similar to what was described above. Overfitting
occurs when a model gets too acquainted with the training data and performs extremely well on the training data in
comparison to the testing data or validation set. This occurs when enough training data isn’t provided (the data set is
too small), where models get too familiar with the noise and unimportant details in the training data and are unable to
identify a general pattern.
A problem that was observed when using a miRNA dataset was the lack of current literature on the topic. The paper
analyzed [69] identified multiple relevant miRNA biomarkers associated with liver cancers. However, because of the little
research that exists on some miRNAs, the study was unable to achieve its potential maximum accuracy. This is because
some important data values were missing. Although miRNA profiling serves as a great means to differentiate and diag-
nose tumors and various diseases, lacking literature creates a strong limitation for using such datasets.
A lack of existing literature also impacted the paper analyzed using a DNA methylation dataset [70]. Although the study
presented with a high accuracy with novel methylation sites associated with HCC, longitudinal studies are required to
accurately confirm the prognostic value of these markers and their link to liver cancer. Lastly, the study analyzed relied
on the use of tumor tissue biopsies to obtain methylation profiles from patients, causing unwanted trauma for diagnosis.
Although newer research has found alternative solutions to this, this paper still was considered unnecessarily invasive,
which challenged its diagnostic efficacy.

6.4 Research investigation 4: suggestions for future research

Certain areas and algorithms were observed to show notable performance and efficacy improvements within the models.
These are discussed below.

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

• The long-lasting problem in the space of liver cancer diagnosis is between PLCs and SLCs. For this, it is vital to consider
the importance of a structured AI model, incorporating model imaging and miRNA tools. For a first diagnosis, imag-
ing is likely the right approach as it has a larger area of view and is easier to interpret. However, when attempting to
differentiate PLCs and SLCs, and when biopsies, or now peripheral blood sampling (utilization of ctDNA) are doable
for miRNA profiling, evidence suggests that these may also be a wise approach. miRNA play a crucial role in gene
expression and often show aberrant patterns in cancer tissue. miRNA profiling can significantly enhance the early
detection and differentiation of primary versus secondary liver tumors. This is because genetic changes are often
more reliable and clearly evident than imaging alone. Future models could incorporate miRNA expression data to
complement imaging data, leading to easier and more efficient early detection. The same applies to a broad range of
genomic-based models, including but not limited to DNA methylation, gene signatures and proteomics. Furthermore,
multi-omics have shown to often be superior to the single-omics models discussed above [85]. showed through their
research that multi-omics models (trained on 3-omic data) showed greater prediction than single-omics-based DL
models. The authors of the study claimed and showed that multi-omics data is better for model training in comparison
to single-omics; however, it is important to acknowledge the difficulty of real-world implementation of multi-omics
data and the difficulty involved in conducting multiple sequencing technologies, attaining patient consent while
avoiding patient stress/risk, and collecting data overall. Also, multi-omics data helps in overcoming the fact that
medical data such as tumor data has high intra-tumoral heterogeneity, where within a single tumor there can be
significant genetic, epigenetic, and phenotypic diversity. Utilizing only one time of data creates limitations as such
models may not capture the full complexity of the tumor.
• When moving towards clinical trials for a deep learning model, it is important to prepare in advance for a broader
testing cohort. Because of this, it is vital to consider rarer cancers, which are often not available in online datasets. A
possible solution to this is to utilize GANs (generative adversarial networks)—these networks are able to generate
synthetic medical images that can then potentially address the issue of a small or lacking dataset [86]. Medical data
helps these networks augment existing datasets and thus improving their performance on underrepresented tumor
types.
• Deep learning image reconstruction (DLIR), a technique utilized by [81, 82] has shown significant implications in the
differentiation between metastatic and primary cancers in the liver. DLIR are a series of deep learning algorithms used
to reduce noise and artifacts and enhance resolution and contrast from imaging modalities like PET and CT scans.
DLIR results in clearer and more detailed medical images. Along with aiding diagnostic precision, it also allows for
better visualization of small lesions that may not be discernible using traditional image reconstruction/enhancement
methods [87]. Moreover, DLIR facilitates more accurate delineation of tumor boundaries, which is critical in diagnosis
and prognostication of this cancer. The enhancements done through DLIR help radiologists distinguish between dif-
ferent tissue types and identify abnormalities with greater accuracy–as shown in [82]. DLIR also produces high quality
and noise reduces images with relatively lower radiation doses, thereby minimizing patient exposure and allowing
for frequent monitoring [88]. Lastly, the clinical implementation of DLIR systems is also relatively more feasible. This
is largely because unlike most, DLIR algorithms are not fully autonomous; instead, they serve as diagnostic aids to
radiologists.
• The usage of curated supervised attention mechanisms is also suggested. A supervised attention mechanism (SAM)
in DL refers to a method of guiding a model’s focus on relevant parts of input data based on external annotations
done by expert radiologists/doctors. This has been shown to enhance AI models’ ability to prioritize relevant features,
significantly improving accuracy [80]. In medical screenings, SAMs allow radiologists to annotate on critical regions
such as tumors or lesions and teach the model to better differentiate through their notes. Schmauch et al. [80] used a
SAM and thereby showed improvement in the model’s ability to differentiate between 5 categories, including HCC and
SLCs. Radiologists manually identified abnormalities and used bounding box annotations when training the model.
This aided in classification as mentioned before, but also improved the interpretability of the model. By overlaying
attention maps on the ultrasound images, the researchers could visualize which areas the model focused on most,
allowing them to understand where the model made errors and where it succeeded–essentially supervising the
model’s diagnosis. SAMs showed in the studies [72, 80] that they have potential to increase accuracy when model’s
are required to classify inputs into non-binary classes (3 or more categories, for example: HCC, ICC, benign, SLC).
• Coming to preprocessing techniques used by the papers analyzed, the combination of Watershed and Gaussian Mix-
ture Model (GMM) stands out by far for imaging datasets. This model showed a 100% sensitivity and 99.38% accuracy.
Watershed transform aids in image segmentation by treating the grayscale image as a topographic surface represent-
ing intensities as peaks, hills and valleys. This tool is used to delineate regions, specifically the liver and tumor regions.

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

By doing this preprocessing, the load and data provided onto the classifier are reduced and the region of interest is
isolated, therefore likely resulting in a higher accuracy. The GM model is used to cluster the pixels in an input image
into different groups based on their intensity values, essentially clearly differentiating tumor and non-tumor regions.
This is further used to isolate the ROI and has been shown to increase both accuracy and sensitivity. Together, these
models show a high potential for clinical use. It is important to note the added computational complexity that results
from these add-ons, however. The main purpose of this is to emphasize the significance of preprocessing and segmen-
tation algorithms in models, not just referring to GMM and Watershed. However, the immensely high accuracy must
be critically evaluated and not taken at face value. These results were achieved with a relatively small dataset of 225 CT
images, which may not be sufficient for robust conclusions. The study used a threefold cross-validation method rather
than the more robust 5 or tenfold. Additionally, the high performances such as this suggest potential for overfitting.
Future studies should validate these findings on larger, more diverse datasets with robust cross-validation methods
to ensure the results are generalizable and reliable in clinical practice.
• Sparsity-based classification may be a solution to the above-mentioned problem of computational complexity. To
address the potential for regular clinical use, it is necessary to maintain computational complexity and user interpret-
ability. This method cuts down on the features being analyzed by identifying and focussing on the most significant
ones. It also has shown promise in its use to reduce false positive rates and increase model efficiency.
• Siamese networks use a twin mechanism, where two identical neural networks are trained in parallel on two differ-
ent inputs. This mechanism is highly useful in differentiating objects of similar types and therefore may be worth
researching in its implications in differentiating between PLCs and between PLCs and SLCs.
• Alongside the previously listed advantages of 3D data and model usage, the 3D architecture proposed in the paper
[68] also removes the need for any preprocessing steps. This refers to cropping, ROI annotating, etc. Although 3D
models and data increase the load put on the model, the removal of such add-ons reduces the complexity in an
attempt to increase efficacy.
• Dynamic and Contrast-enhanced imaging showed higher accuracies in the analyzed studies, likely due to the added
data with dynamic images and easier delineation and segmentation through contrast.
• As mentioned in the previous section, dataset handling is also vital to create generalizable models that are practical
to use in real-world scenarios. Medical datasets commonly consist of class imbalance, where different instances of
each class are not equally distributed. GANs mentioned previously are one approach to this problem; however, we
also emphasize the importance of oversampling and undersampling techniques for class balancing. This has been
elaborated in Sect. 5.
• Lastly, we also want to emphasize the astonishing growth and implementation of transfer learning in the field of
medical images. As detailed thoroughly in Sect. 3.3, transfer learning modes aid in reducing manual interaction, deline-
ation, segmentation, and can be converted into hybrid models enabling them with higher precision and accuracies.
We encourage further investigations into individual and hybrid transfer learning models.

We propose to evaluate the performance of the proposed machine learning approach on multiple datasets represent-
ing varied patient demographics. Also, we acknowledge the need for further validation in real-world clinical environments
for assessing the generalizability. In the future, we aim to address all the issues raised in the study regarding previous
research and focus on testing the model with real-world datasets in distinct clinical settings. This can assure the model’s
generalizability, thus enhancing its clinical utility and reliability [89, 90].

6.4.1 Clinical impacts of AI‑based predictive models

• IBM Watson oncology: Oncologists can use IBM Watson for Oncology to make better treatment decisions. Watson
helps hospitals globally, like as Manipal Hospitals in India, personalize cancer treatment programs by examining a
patient’s medical history, genetics, and the latest research [91].
• Google AI for diabetic retinopathy: Google AI’s deep learning algorithm can detect diabetic retinopathy, which can
cause blindness if untreated. Indian clinics use this technique to swiftly and precisely determine at-risk patients. This
AI solution prioritizes urgent patients, decreasing specialist burden and assuring prompt interventions [92].
• Aidoc for radiology: Aidoc helps radiologists discover medical picture anomalies with AI. It helps radiologists prioritize
critical cases like CT scan strokes and cerebral hemorrhages [93].

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

• PathAI for pathology: Machine learning tools from PathAI help pathologists diagnose diseases from tissue samples.
It increases pathology results consistency and reduces diagnosis errors by giving pathologists AI-powered insights
[94].

7 Conclusion

Liver cancer is a highly prevalent form of cancer, with nearly 800,000 new cases annually, and this number is expected to
increase over time. However, studies have shown that artificial intelligence has made significant progress in tackling this
issue by creating predictive DL models that have shown good accuracy and performance for diagnosing liver cancers.
Despite these advancements, our research identifies several gaps that need to be addressed to enhance the efficacy of
predictive models. A critical weakness identified across many studies is the inability of AI-based models to efficiently
differentiate between PLCs and SLCs. Additionally, underutilization of 3D datasets, problems with dataset origin, dataset
type, and more have also been identified as areas that need to be addressed in future research. This investigative review
answered key questions and the results are summed as follows: We detail the evaluation parameters that should ideally
be used to evaluate performance of future predictive models. We also explain common liver cancer datasets, detailing
various types, and provide a comprehensive list of research gaps in current literature. Lastly, we have also listed potential
suggestions for future research, backed by evidence from our research. This review can guide the development of future
models by providing crucial information about valuable algorithms, preprocessing techniques, and datasets that can
be used to achieve high performance in AI-based models in the future. These insights, along with previous research, will
help AI models improve as efficient decision-support systems for doctors in the field of cancer diagnosis.

Author contributions S. Grover: Conceptualization, Writing – original draft, Data Collection, Critical Review S. Gupta: Writing – review & edit-
ing, Supervising.

Funding The authors have not disclosed any funding.

Data availability The datasets presented in this study can be found in online repositories. The names of the repository/repositories and acces-
sion number(s) can be found in the article.

Declarations
Competing interests The authors declare that the research was conducted in the absence of any commercial or financial relationships that
could be construed as a potential conflict of interest.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which
permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to
the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You
do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party
material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds
the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://​creat​iveco​
mmons.​org/​licen​ses/​by-​nc-​nd/4.​0/.

References
1. Hanahan D, Weinberg RA. The hallmarks of cancer. Cell. 2000;100(1):57–70.
2. Lengauer C, Kinzler KW, Vogelstein B. Genetic instabilities in human cancers. Nature. 1998;396(6712):643–9.
3. Wu K, El Zowalaty AE, Sayin VI, Papagiannakopoulos T. The pleiotropic functions of reactive oxygen species in cancer. Nat Cancer.
2024;5(3):384–99.
4. Bray F, et al. The ever-increasing importance of cancer as a leading cause of premature death worldwide. Cancer. 2021;127(16):3029–30.
5. Celià-Terrassa T, Kang Y. How important is EMT for cancer metastasis? PLoS Biol. 2024;22(2): e3002487.
6. Abbasnia V, Foadoddini M, Esfahani DE, Khazdair MR, Oryan S. Rosmarinic acid improves tracheal smooth muscle responsiveness and
lung pathological changes in ovalbumin-sensitized rats. Asian Pac J Trop Biomed. 2024;14(3):100–5.
7. Kader T, Hill P, Rakha EA, Campbell IG, Gorringe KL. Atypical ductal hyperplasia: update on diagnosis, management, and molecular land-
scape. Breast Cancer Res. 2018;20:1–11.

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

8. Chou CT, Chou JM, Chang TA, Huang SF, Chen CB, Chen YL, Chen RC. Differentiation between dysplastic nodule and early-stage hepato-
cellular carcinoma: the utility of conventional MR imaging. World J Gastroenterol. 2013;19(42):7433–9.
9. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of inci-
dence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clinicians. 2024;74(3):229–63.
10. Chen Q, Rui F, Ni W, Li J. Research progress in epidemiology and risk factors of primary liver cancer. Chin Gen Pract. 2024;27(06):637.
11. Li Q, Cao M, Lei L, Yang F, Li H, Yan X, He S, Zhang S, Teng Y, Xia C, Chen W. Burden of liver cancer: from epidemiology to prevention. Chin
J Cancer Res = Chung-kuo Yen Cheng Yen Chiu. 2022;34(6):554–66.
12. Susanti C, Mariska A. Utilization of biopsy-guided CT scan in diagnosing liver cancer: a case study. Sriwijaya J Radiol Imaging Res.
2024;2(1):50–5.
13. Massarweh NN, El-Serag HB. Epidemiology of hepatocellular carcinoma and intrahepatic cholangiocarcinoma. Cancer Control.
2017;24(3):1073274817729245.
14. Park JH, Kim JH. Pathologic differential diagnosis of metastatic carcinoma in the liver. Clin Mol Hepatol. 2019;25(1):12–20.
15. Staniczek T. Endothelial regulation of liver homeostasis and regeneration (Doctoral dissertation); 2024.
16. Obenauf AC, Massagué J. Surviving at a distance: organ-specific metastasis. Trends Cancer. 2015;1(1):76–91.
17. Cohen R, Raeisi M, Chibaudel B, Shi Q, Yoshino T, Zalcberg JR, de Gramont A. Prognostic value of liver metastases in colorectal cancer
treated by systemic therapy: An ARCAD pooled analysis. Eur J Cancer. 2024;207:114160.
18. Rumgay H, Arnold M, Ferlay J, Lesi O, Cabasag CJ, Vignat J, Laversanne M, McGlynn KA, Soerjomataram I. Global burden of primary liver
cancer in 2020 and predictions to 2040. J Hepatol. 2022;77(6):1598–606.
19. Gonzalez SA, Keeffe EB. Diagnosis of hepatocellular carcinoma: role of tumor markers and liver biopsy. Clin Liver Dis. 2011;15(2):297–306.
20. Görgec B, Hansen IS, Kemmerich G, Syversveen T, Hilal MA, Belt EJ, Zonderhuis BM. MRI in addition to CT in patients scheduled for local
therapy of colorectal liver metastases (CAMINO): an international, multicentre, prospective, diagnostic accuracy trial. Lancet Oncol.
2024;25(1):137–46.
21. Lindemann ME, Gratz M, Grafe H, Jannusch K, Umutlu L, Quick HH. Systematic evaluation of human soft tissue attenuation correction
in whole-body PET/MR: Implications from PET/CT for optimization of MR-based AC in patients with normal lung tissue. Med Phys.
2024;51(1):192–208.
22. Chmielewski M, Serafin Z, Kamińska D, Skrobisz K, Kozak O, Olczyk P, Krajewska M. The use of intravascular contrast media in patients with
impaired kidney function–joint clinical practice position statement of the polish society of nephrology and the polish medical society
of radiology. Pol J Radiol. 2024;89:e161.
23. Hameed U, Ur Rehman M, Rehman A, Damaševičius R, Sattar A, Saba T. A deep learning approach for liver cancer detection in CT scans.
Comput Methods Biomech Biomed Eng Imaging Vis. 2024;11(7):2280558.
24. Sabirovna IN, Raykhona K. Benefits of ultrasound examination. Tadqiqotlar. 2024;31(2):95–100.
25. Barba M, Cola A, De Vicari D, Costa C, La Greca G, Vigna A, Maruccia S. Changes in pelvic floor ultrasonographic features after flat magnetic
stimulation in women with chronic pelvic pain and levator ani muscle hypertonicity. Medicina. 2024;60(3):374.
26. Gadour E, Awad A, Hassan Z, Shrwani KJ, Miutescu B, Okasha HH. Diagnostic and therapeutic role of endoscopic ultrasound in liver dis-
eases: a systematic review and meta-analysis. World J Gastroenterol. 2024;30(7):742–58.
27. Radetic M, DeVita R, Haaga J. When is contrast needed for abdominal and pelvic CT? Clevel Clin J Med. 2020;87(10):595–8.
28. O’Connor JP, Tofts PS, Miles KA, Parkes LM, Thompson G, Jackson A. Dynamic contrast-enhanced imaging techniques: CT and MRI. Br J
Radiol. 2011;84(Special_Issue_2):S112–20.
29. Hylton N. Dynamic contrast-enhanced magnetic resonance imaging as an imaging biomarker. J Clin Oncol. 2006;24(20):3293–8.
30. Macfarlane LA, Murphy PR. MicroRNA: biogenesis, function and role in cancer. Curr Genomics. 2010;11(7):537–61.
31. Shademan B, Karamad V, Nourazarian A, Masjedi S, Isazadeh A, Sogutlu F, Avcı CB. MicroRNAs as targets for cancer diagnosis: interests
and limitations. Adv Pharm Bull. 2023;13(3):435–45.
32. Valinezhad Orang A, Safaralizadeh R, Kazemzadeh-Bavili M. Mechanisms of miRNA-mediated gene regulation from common downregula-
tion to mRNA-specific upregulation. Int J Genom. 2014;2014: 970607.
33. O’Brien J, Hayder H, Zayed Y, Peng C. Overview of MicroRNA biogenesis, mechanisms of actions, and circulation. Front Endocrinol.
2018;9:402.
34. Chakrabortty A, Patton DJ, Smith BF, Agarwal P. miRNAs: potential as biomarkers and therapeutic targets for cancer. Genes. 2023;14(7):1375.
35. Wang J, Chen J, Sen S. MicroRNA as biomarkers and diagnostics. J Cell Physiol. 2016;231(1):25–30.
36. Moore LD, Le T, Fan G. DNA methylation and its basic function. Neuropsychopharmacology. 2013;38(1):23–38.
37. Dong Y, Zhao H, Li H, Li X, Yang S. DNA methylation as an early diagnostic marker of cancer. Biomed Rep. 2014;2(3):326–30.
38. Draškovič T, Hauptman N. Discovery of novel DNA methylation biomarker panels for the diagnosis and differentiation between common
adenocarcinomas and their liver metastases. Sci Rep. 2024;14(1):3095.
39. Yang M, Park JY. DNA methylation in promoter region as biomarkers in prostate cancer. Methods Mol Biol (Clifton, NJ). 2012;863:67–109.
40. Zhang C, Sheng Q, Zhao N, Huang S, Zhao Y. DNA hypomethylation mediates immune response in pan-cancer. Epigenetics.
2023;18(1):2192894.
41. Singer BD. A practical guide to the measurement and analysis of DNA methylation. Am J Respir Cell Mol Biol. 2019;61(4):417–28.
42. Das PM, Singal R. DNA methylation and cancer. J Clin Oncol. 2004;22(22):4632–42.
43. Tandon R, Agrawal S, Rathore NPS, Mishra AK, Jain SK. A systematic review on deep learning-based automated cancer diagnosis models.
J Cell Mol Med. 2024;28(6): e18144.
44. Mao WB, Lyu JY, Vaishnani DK, Lyu YM, Gong W, Xue XL, Shentu YP, Ma J. Application of artificial neural networks in detection and diagnosis
of gastrointestinal and liver tumors. World J Clin Cases. 2020;8(18):3971–7.
45. Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional neural networks: an overview and application in radiology. Insights Imaging.
2018;9(4):611–29.
46. Asiri N, Hussain M, Al Adel F, Alzaidi N. Deep learning based computer-aided diagnosis systems for diabetic retinopathy: a survey. Artif
Intell Med. 2019;99: 101701.
47. Hao S, Zhou Y, Guo Y. A brief survey on semantic segmentation with deep learning. Neurocomputing. 2020;406:302–21.

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0 Review

48. Noble W. What is a support vector machine? Nat Biotechnol. 2006;24:1565–7.

49. Lai T. Interpretable medical imagery diagnosis with self-attentive transformers: a review of explainable AI for health care. BioMedInfor-
matics. 2024;4(1):113–26.
50. Valanarasu JMJ, Oza P, Hacihaliloglu I, Patel VM. Medical transformer: Gated axial-attention for medical image segmentation. In: Medi-
cal image computing and computer assisted intervention–MICCAI 2021: 24th international conference, Strasbourg, France, September
27–October 1, 2021, proceedings, part I 24. Springer International Publishing; 2021. pp. 36–46.
51. Qi Q, Lin L, Zhang R, Xue C. MEDT: Using multimodal encoding-decoding network as in transformer for multimodal sentiment analysis.
IEEE Access. 2022;10:28750–9.
52. Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. In: Medical image computing and
computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5–9, 2015, proceedings, part
III 18. Springer International Publishing; 2015. pp. 234–241.
53. Chen J, Lu Y, Yu Q, Luo X, Adeli E, Wang Y, Zhou Y. Transunet: transformers make strong encoders for medical image segmentation; 2021.
arXiv preprint arXiv:2102.04306.
54. Ayalew YA, Fante KA, Mohammed MA. Modified U-Net for liver cancer segmentation from computed tomography images with a new
class balancing method. BMC Biomed Eng. 2021;3:1–13.
55. Hasan DF, Khidhir AM. Toward enhancement of deep learning techniques using fuzzy logic: a survey. Int J Electr Comput Eng (IJECE).
2023;13(3):3041–55.
56. Kumar D, Singh J, Singh OP. A fuzzy logic based decision support system for evaluation of suppliers in supply chain management
practices. Math Comput Model. 2013;58(11–12):1679–95.
57. Lin CT, Lee CSG. Neural-network-based fuzzy logic control and decision system. IEEE Trans Comput. 1991;40(12):1320–36.
58. Sun J, Qi C, Liu Y, Gao F, Fu X, Tian Y. Evaluation of multiple liver cancer scoring systems. Adv Biol. 2024;8(2):2300301.
59. Summerlin DS, Porter KK. Differentiating focal liver lesions: ICC and HCC. Appl Radiol. 2022;51(2):36–8.
60. Zhen SH, Cheng M, Tao YB, Wang YF, Juengpanich S, Jiang ZY, Jiang YK, Yan YY, Lu W, Lue JM, Qian JH, Wu ZY, Sun JH, Lin H, Cai XJ. Deep
learning for accurate diagnosis of liver tumor based on magnetic resonance imaging and clinical data. Front Oncol. 2020;10:680.
61. Survarachakan S, Prasad PJR, Naseem R, de Frutos JP, Kumar RP, Langø T, Lindseth F. Deep learning for image-based liver analysis—a
comprehensive review focusing on malignant lesions. Artif Intell Med. 2022;130:102331.
62. Low RN, Semelka RC, Worawattanakul S, Alzate GD. Extrahepatic abdominal imaging in patients with malignancy: comparison of MR
imaging and helical CT in 164 patients. J Magn Reson Imaging Off J Int Soc Magn Reson Med. 2000;12(2):269–77.
63. Ahmad G, Khan MA, Abbas S, Athar A, Khan BS, Aslam MS. Automated diagnosis of hepatitis B using multilayer mamdani fuzzy infer-
ence system. J Healthc Eng. 2019;2019:6361318.
64. Alkhalifah T, Singla J, Alurise F. Fuzzy logic-based system for liver fibrosis disease. Comput Syst Sci Eng. 2023;46(3):3559–82.
65. Mirmozaffari M. Developing an expert system for diagnosing liver diseases. Eur J Eng Technol Res. 2019;4(3):1–5.
66. Ekong VE, Onibere EA, Imianvan AA. Fuzzy cluster means system for the diagnosis of liver diseases. Int J Comput Sci Technol.
2011;2(3):205–9.
67. Yasaka K, Akai H, Abe O, Kiryu S. Deep learning with convolutional neural network for differentiation of liver masses at dynamic
contrast-enhanced CT: a preliminary study. Radiology. 2018;286(3):887–96.
68. Trivizakis E, Manikis GC, Nikiforaki K, Drevelegas K, Constantinides M, Drevelegas A, Marias K. Extending 2-D convolutional neural
networks to 3-D for advancing deep learning cancer classification with application to MRI liver tumor differentiation. IEEE J Biomed
Health Inform. 2018;23(3):923–30.
69. Sayed GI, Solyman M, El Gedawy G, Moemen YS, Aboul-Ella H, Hassanien AE. Circulating miRNA’s biomarkers for early detection of
hepatocellular carcinoma in Egyptian patients based on machine learning algorithms. Sci Rep. 2024;14(1):4989.
70. Bai Y, Tong W, Xie F, Zhu L, Wu H, Shi R, Zhang Y. DNA methylation biomarkers for diagnosis of primary liver cancer and distinguishing
hepatocellular carcinoma from intrahepatic cholangiocarcinoma. Aging (Albany NY). 2021;13(13):17592.
71. Gao R, Zhao S, Aishanjiang K, Cai H, Wei T, Zhang Y, Gu J. Deep learning for differential diagnosis of malignant hepatic tumors based
on multi-phase contrast-enhanced CT and clinical data. J Hematol Oncol. 2021;14:1–7.
72. Wang Q, Wang Z, Sun Y, Zhang X, Li W, Ge Y, Chen Y. SCCNN: A diagnosis method for hepatocellular carcinoma and intrahepatic
cholangiocarcinoma based on siamese cross contrast neural network. IEEE Access. 2020;8:85271–83.
73. Yang Y, Zhou Y, Zhou C, Ma X. Deep learning radiomics based on contrast enhanced computed tomography predicts microvascular
invasion and survival outcome in early stage hepatocellular carcinoma. Eur J Surg Oncol. 2022;48(5):1068–77.
74. Das A, Acharya UR, Panda SS, Sabut S. Deep learning based liver cancer detection using watershed transform and Gaussian mixture
model techniques. Cogn Syst Res. 2019;54:165–75.
75. Rajkumar TD, Deepa D, Jeyaranjani J. Automatic diagnosis of liver tumor in CT images. Int J Eng Adv Technol. 2019;9(1s4):1105–9.
76. Chlebus G, Schenk A, Moltz JH, van Ginneken B, Hahn HK, Meine H. Automatic liver tumor segmentation in CT with fully convolutional
neural networks and object-based postprocessing. Sci Rep. 2018;8(1):15497.
77. Ben-Cohen A, Klang E, Kerpel A, Konen E, Amitai MM, Greenspan H. Fully convolutional network and sparsity-based dictionary learn-
ing for liver lesion detection in CT examinations. Neurocomputing. 2018;275:1585–94.
78. Liu X, et al. Identification of multi-omics biomarkers and construction of the novel prognostic model for hepatocellular carcinoma.
Sci Rep. 2022;12(1):12084.
79. Lu X, et al. Identification of key genes in hepatocellular carcinoma and validation of the candidate gene, cdc25a, using gene set
enrichment analysis, meta-analysis and cross-species comparison. Mol Med Rep. 2016;13(2):1172–8.
80. Schmauch B, et al. Diagnosis of focal liver lesions from ultrasound using deep learning. Diagn Interv Imaging. 2019;100(4):227–33.
81. Suyama J, et al. Diagnostic value of deep learning image reconstruction for detection of liver metastases on FDG-PET with digital
PET: comparison with EOB-MRI. JNM. 2023;64:P1022–P1022.
82. Kanan A, et al. Deep learning CT reconstruction improves liver metastases detection. Insights Imaging. 2024;15(1):167.
83. Shen J, et al. Identification of a novel gene signature for the prediction of recurrence in HCC patients by machine learning of genome-
wide databases. Sci Rep. 2020;10(1):4435.

Vol.:(0123456789)
Review Discover Applied Sciences (2024) 6:508 | https://doi.org/10.1007/s42452-024-06218-0

84. Wei Q, Tan N, Xiong S, Luo W, Xia H, Luo B. Deep learning methods in medical image-based hepatocellular carcinoma diagnosis: a
systematic review and meta-analysis. Cancers. 2023;15(23):5701.
85. Chaudhary K, et al. Deep learning–based multi-omics integration robustly predicts survival in liver cancer. Clin Cancer Res.
2018;24(6):1248–59.
86. Abdusalomov AB, Nasimov R, Nasimova N, Muminov B, Whangbo TK. Evaluating synthetic medical images using artificial intelligence
with the GAN algorithm. Sensors. 2023;23(7):3440.
87. Cao L, et al. Improving spatial resolution and diagnostic confidence with thinner slice and deep learning image reconstruction in
contrast-enhanced abdominal CT. Eur Radiol. 2023;33(3):1603–11.
88. Chandran MO, Pendem S, Priya PS, Chacko C, Kadavigere R. Influence of deep learning image reconstruction algorithm for reducing
radiation dose and image noise compared to iterative reconstruction and filtered back projection for head and chest computed
tomography examinations: a systematic review. F1000Research. 2024;13:274.
89. Nissar I, Alam S, Masood S, Mir WA. Bridging the gap between technology and medicine: approaches of artificial intelligence in
healthcare. In: Machine learning and artificial intelligence in healthcare systems. CRC Press; 2023. pp. 173–190.
90. Li S, Li X, Yu K, Miao D, Zhu M, Yan M, Liu N. Bridging data gaps in healthcare: a scoping review of transfer learning in biomedical data
analysis; 2024. arXiv preprint arXiv:2407.11034.
91. Tucker T. IBM watson health: managing partnerships to improve personalized medicine. Cases and Tools in Biotechnology Management.
92. Bellemo V, Lim G, Rim TH, Tan GS, Cheung CY, Sadda S, Ting DSW. Artificial intelligence screening for diabetic retinopathy: the real-world
emerging application. Curr Diabet Rep. 2019;19:1–12.
93. Stephens K (2021) Radiology partners, aidoc partner to accelerate the use of artificial intelligence. AXIS Imaging News.
94. Baxi V, Edwards R, Montalto M, Saha S. Digital pathology and artificial intelligence in translational medicine and clinical practice. Mod
Pathol. 2022;35(1):23–32.
95. Ramani A, Tapper EB, Griffin C, Shankar N, Parikh ND, Asrani SK. Hepatocellular carcinoma-related mortality in the USA, 1999–2018. Dig
Dis Sci. 2022;67(8):4100–11.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Vol:.(1234567890)