Image Denoising Using Auto-Encoder
Group Members:- Guide:-
1.Kuldeep Gangwar (2107340130030) Mr.Abhishek Kr. Yadav
2.Rishi Muni (2107340130056) Ms.Neha Sankhwar
3.Siddharth Singh (2107340130061)
4.Sonu Prajapati (2107340130063)
Introduction
Medical imaging techniques such as X-Rays,MRIs,CT-scans & Ultrasounds often suffers from noise
during acquisition due to limitations in equipment ,environmental factors or patient movement.This
noise can obscure critical anatomical details, making difficulty to detect accurately diagnose for
healthcare specialist.
Traditional denoising method such as spatial filtering method to fails or either remove complex noise
patterns ,as a result there is a need for advanced denoising techniques using autoencoder that can
enhance image quality without significant loss of information.
Problem Statement
To remove noise from medical images, while preserving critical details, improving diagnosis
accuracy by learning efficient representations for clean reconstruction.
Objective
The objective of image denoising using auto-encoder in field of medical imaging is to improve the
quality of medical images.
1. Reduce Poission Noise:-This noise becomes more pronounced in darker or less intense regions of image.
Poission noise arises in medical imaging systems ,particularly low dose modalities like CT scans.
2. Speckle noise (multiplying random pixel values with different pixels of an image) :- This type noise
commonly seen in ultrasound and MRI imaging.
3. Improve Diagnostic Accuracy: -Enhance clarity &visibility of anatomical structure.
4. Preserve Essential Features:-Retain fine details & important features of image.
Methodology
Adding Noise to Images :
To simulate real-world conditions and evaluate the denoising model's effectiveness, noise is added to clean
medical images. This step is crucial for training the autoencoder model. Common types of noise added include:
Poisson Noise: Often used for images that represent photon counting or intensity-based medical images.
Gaussian Noise: Random noise added to pixel values following a Gaussian distribution. This is useful for
simulating sensor noise.
Autoencoders
An autoencoder is a type of artificial neural network used to learn efficient coding's of unlabelled
data (unsupervised learning). An autoencoder learns two functions: an encoding function that transforms
the input data, and a decoding function that recreates the input data from the encoded representation. The
architecture of an autoencoder typically consists of two main components:
Encoder: The encoder comprises layers that encode a compressed representation of the input data
through dimensionality reduction.
• Decoder: The decoder comprises hidden layers with a progressively larger number of nodes that
decompress (decode) the encoded representation of data, ultimately reconstructing the data back to
its original, pre-encoding form.
Neural Network Layers
The autoencoder model used for medical image denoising typically consists of several key
neural network layers:
Result evaluation parameters
PSNR: Peak Signal to Noise Ratio is unit for measuring the amount of noise in an image. Its value is
number of clear pixels per each noisy pixel in an image.
MSE: Mean Square Error is the average squared difference between the es timated value and the actual
value also known as loss.
LITRATURE REVIEW
S. Ahmed et al. [1] In this paper the work is carried on Stacked Convolutional Autoencoder (SCAE)- based
denoising system has proven to be highly effective in enhancing the quality of medical images, particularly in
reducing noise in challenging datasets such as 2-DGE images. This novel approach to medical image denoising
demonstrates the potential of deep learning-based models for improving image quality, particularly in applications
like proteomics and biomarker discovery, where image clarity is crucial for accurate analysis.
K. Bajaja et al. [2] Proposes a deep convolutional denoising autoencoder framework to remove Gaussian noise,
utilizing skip connections to address gradient fading and optimize image restoration. Experimental results show
superior PSNR performance compared to traditional methods. Future work includes addressing other noise types
and applications like image inpainting, super-resolution, and sound/video denoising.
Pawar [3] presents an approach to reduce image noise using image processing and deep learning techniques,
specifically autoencoders. Unlike traditional methods, autoencoders focus on preserving the originality of the
signal through backpropagation. The proposed technique is versatile, reliable, efficient, and compatible with a
wide range of devices. proposes a noise removal technique using an autoencoder. The autoencoder compresses
input images into a Latent Space Representation (LSR) vector during encoding, then reconstructs the image
during decoding. The loss of certain features during decoding helps remove noise, as these lost features are
primarily noise. The altered dataset, initially containing noisy images, is processed by the autoencoder to
effectively denoise the signal.
L. Gondara [4] shown that convolutional denoising autoencoders can effectively clean medical images, even
with small training datasets—just 300 samples are enough for good results. In the future, we plan to explore the
best network architecture for small datasets, test it on high-resolution images, and combine other denoising
methods like SVD and median filters. We also aim to see if using additional datasets, like ImageNet, can improve
performance by increasing the number of training samples.
Mizuho Nishio [5] This study validated a patch-based denoising method for ultra-low-dose CT (ULD-CT)
images using a convolutional auto-encoder neural network and image pairs of standard-dose CT and ULD-CT.
Chest phantom images were acquired at 300 mA and 10 mA. Radiologists and technologists evaluated the
denoised images for artifacts, noise, pulmonary vessel visualization, and overall quality, demonstrating the
method's effectiveness. This paper proposes using a convolutional auto-encoder (CAE) to denoise ultra-low-
dose CT (ULDCT) images. While neural networks have effectively denoised non-medical images, studies on
medical image denoising primarily use conventional neural networks. Literature indicates limited application of
deep learning, such as CAE, for CT image enhancement, with only a few studies demonstrating its
effectiveness.
Implementation :
