ONLINE NOTES APP

APROJECTREPORT
Submittedinpartialfulfillmentoftherequirementsfor
theawardof
BACHELOR OFTECHNOLOGY
In
ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

SubmittedBy

R.v.prasannakumar:22A21A6193
Surya:22A21A6198
Tarun:22A21A61B5
Narasimha:22A21A6178
Sravani:22A21A6195
Underthe EsteemedGuidanceof
CH.Chandrika Surya
M.Tech
AssociateProfessor

DEPARTMENT OF ARTIFICIAL INTELLIGENCE AND MACHINE

LEARNING
SWARNANDHRACOLLEGEOFENGINEERING&TECHNOLOGY
(ApprovedbyAICTE, AffiliatedtoJNTU-Kakinada,AccreditedbyNAAC)

(AUTONOMOUS)
Seetharampuram, Narsapur–534280,W.G.Dt.(A.P)

2024-2025
 SWARNANDHRACOLLEGEOFENGINEERING&TECHN
OLOGY
(Approved By AICTE &Affliated To JNTU-Kakinada,Accredited by NAAC)

(AUTONOMOUS)

SEETHARAMAPURAM, NARSAPUR-534280,W.G.Dt.(A.P)

DEPARTMENTOFARTIFICIAL INTELLIGENCE AND MACHINE

LEARNING

Certificate

Certified that this project work titled

“CALCULATOR”isabonafidework of R.v.prasannakumar-
22A21A6193,Surya-22A21A6198,Tarun-22A21A61B5,Narasimha-
22A21A6178,Sravani-22A21A6195,of 3rd B.Tech who carried out the
workunder my supervision, and submitted in partial fulfillment of the
requirements fortheawardofdegree,BACHELOR OFTECHNOLOGYin
ARTIFICIAL INTELLIGENCE AND MACHINE
LEARNING,duringtheacademicyear2024-2025.

Guide Headof theDepartment

Mrs.CH.Chandrika Surya Dr.B.Rama Krishna
AssociateProfessor Professor

EXTERNALEXAMINER
 ACKNOWLEDGEMENT
Weextendourheartfeltgratitudetothealmightyforgiving
me strength in proceeding with this project. With profound gratitude,respect
and pride, we express our sincere thanks to Management
Members,secretary and correspondent of our college for making necessary
arrangementfordoingtheproject.
WewouldliketothankDr.S.Suresh Kumar,principal forhistimely
suggestions and for giving us permission to carry out the project.
WeWouldliketoexpressourgratefulthankstoDr.B.Rama Krishna,HOD.AIML
DEPT,forhisvaluablesuggestionsandguidanceinregardingthesoftwareanalysis,
DesignandalsoforhiscontinuouseffortinsuccessfulcompletionoftheProject.
Our deep gratitude to internal guide Mrs.CH.Chandrika Surya.
We thankher for dedication, guidance, council and keen interest at every stage
of theproject.
Finallywethankoneandallthathavecontributeddirectlyorindirectlytot
hisProject.

R.v.prasannakumar:22A21A6193
Surya:22A21A6198
Tarun:22A21A61B5
Narasimha:22A21A6178
Sravani-22A21A6195

DECLARATION
 Wecertifythat

a.Theprojectworkcontainedinthethesisisoriginalandhasbeendonebyunde
rtheguidanceofmy supervisor.

b.Theworkhasnotbeensubmittedtoanyother
universityfortheawardofanydegreeor diploma.

c.Theguidelinesoftheuniversityarefollowedinwritingthethesis.

Date:

Place:

Regd.no Name of the student signature

22A21A6193 R.v.prasannakumar
22A21A6198 Surya
22A21A61B5 Tarun
22A21A6178 Narasimha
22A21A6195 Sravani

ABSTRACT
The project is an online notes application built using HTML, CSS, and
JavaScript, designed to help users manage personal or professional notes
efficiently. The core purpose of the app is to provide an intuitive interface where
users can create, update, and organize their notes in one place. The use of HTML
ensures the structure and functionality of the app, while CSS is responsible for the
visual design, ensuring that the app remains user-friendly and visually appealing.
JavaScript is employed to add interactivity and manage dynamic features like
note creation, deletion, and modification.

One of the key features of this notes app is the ability to add new notes. Users can
input text directly into the interface and save it, making the process of note-taking
quick and convenient. The app ensures that each note is stored in a structured
format, allowing for easy retrieval and management. This feature is crucial for
users who want to keep track of various tasks, thoughts, or information, all in a
single digital environment.

Another essential feature is the capability to delete notes. Users can remove
unnecessary or outdated notes with a simple action, ensuring that the interface
remains clutter-free and organized. This adds flexibility to the app, allowing users
to manage their notes effectively, without the need for manual cleanup. The delete
function is seamlessly integrated into the interface, offering both speed and
efficiency for a smoother user experience

The ability to modify existing notes further enhances the app's usability. Users
can edit the content of their notes, making it easier to update information or
correct mistakes. This feature ensures that notes are always relevant and up to
date, contributing to better note management. Additionally, the app automatically
logs the date when a note is created or modified, allowing users to track changes
and updates over time.

Lastly, the app includes a file upload feature, enabling users to attach relevant
documents or images to their notes. This expands the functionality of the app,
making it more than just a simple note-taking tool. By allowing file uploads, the
app becomes a comprehensive platform for organizing information, supporting
multimedia content, and providing users with a complete solution for managing
both text-based and file-based data
CHAPTER TOPIC PAGE NO.
1 INTRODUCTION
1.1 Structure and Function of the QuizApplication

1.2 Project Overview and Steps Involved

1.3 Project Motivation

1.4 Goals and Objectives

Key Features and Benefits

1.5
2 System Architecture
2.1 Components

2.2 Data Flow

2.3 Technology Stack

2.4 System Diagram

3 User Interface Design

3.1 User Interface Elements

3.2 User Flow

4 Quiz Creation and Management

4.1 Creating Quizzes

4.2 Editing Quizzes

4.4 Managing Questions and Answers

5 User Interaction
5.1 Taking Quizzes

5.2 Viewing Results

5.3 User Feedback

5.4 User Profile and Settings

6 Administrator Features
 6.1 Admin Login and Access

6.2 User Management

6.3 Quiz Statistics and Reports

6.4 System Maintenance

7 Technology Implementation
7.1 Programming Languages and Frameworks

7.2 Database Design

7.3 APIs and Integrations

8 Testing and Quality Assurance

8.1 Integration Testing

8.2 User Acceptance Testing

9 Conclusion

10 References
 1. INTRODUCTION
The online notes app is a web-based tool designed to enhance the way users manage and
organize their notes. Built using HTML, CSS, and JavaScript, this application offers a
seamless experience for creating, modifying, and deleting notes, all within an intuitive
and user-friendly interface. In today’s fast-paced world, effective note-taking and
organization are essential for both personal and professional productivity. This app
addresses these needs by offering a comprehensive platform that not only supports text-
based notes but also allows users to upload files and track modifications with
timestamps.

The primary goal of this app is to simplify the note-taking process while providing the
flexibility to manage notes efficiently. Whether it's jotting down ideas, managing to-do
lists, or storing important information, the app ensures that everything is readily
available and easily accessible. Its sleek design, powered by CSS, makes it visually
appealing, while the JavaScript functionality ensures that the app responds quickly to
user actions, making note management a breeze.

Moreover, the inclusion of file uploads and the ability to track creation and modification
dates make this app a versatile tool. It goes beyond just capturing thoughts, allowing users
to attach relevant documents or images, making it a powerful solution for a wide range of
note-taking needs.

1.1 Structure and Function of the Quiz Application

1. HTML (HyperText Markup Language): The structure of the app is built using
HTML, which defines the elements and layout of the interface. Key sections include:
o Header: Contains the app title or logo.
o Main Area: Where users interact with the notes. It includes an input field for creating
new notes, a list area for displaying all existing notes, and buttons for actions like edit,
delete, and file upload.
o Footer: Displays additional information, such as the date and time when a note is
created or modified.
2. CSS (Cascading Style Sheets): CSS is used to style the app, ensuring a clean, user-
friendly interface. Key styling elements include:
o Layout and Responsiveness: Ensures the app looks good on all screen sizes (desktop,
tablet, mobile).
o Button and Input Design: Creates visually appealing buttons for adding, editing, and
deleting notes.
o Note Display: Designs how notes are presented in the list, ensuring readability and
organization.
3. JavaScript: JavaScript handles the interactive and dynamic aspects of the app. It
powers the core functionalities such as creating, modifying, and deleting notes, as well
as file uploads and date tracking. JavaScript also ensures the smooth functioning of user
interactions, ensuring data updates in real time.

Functions
1. Add New Notes:
o Users can add notes by typing in the input field and clicking a "Create Note" button.
o The note is immediately saved and displayed in the list of existing notes.
o Each note is saved with a timestamp showing the creation date and time, which can be
used for tracking purposes.
 2. Delete Notes:
o A delete button is available next to each note, allowing users to remove unwanted or
outdated notes.
o This function is powered by JavaScript and ensures that the notes are removed instantly
from the interface.
3. Modify Notes:
o Users can edit existing notes by clicking an "Edit" button.
o Once modified, the note is updated in the display, and the modification date and time
are automatically logged, ensuring users know when the last changes were made.
4. File Upload:
o Users can upload files (such as images or documents) and attach them to specific notes.
o This feature allows for more detailed note-taking and expands the functionality of the
app to support multimedia content.
5. Date Tracking:
o Each note has a recorded creation date, and if modified, the modification date is
updated.
o This helps users track when notes were added and when they were last edited, offering
better organization and historical context for their notes.

1.2Project Overview and Steps Involved

Project Overview
The online notes app is a web-based tool designed to streamline note-taking and organization
through a user-friendly interface. Built using HTML for structure, CSS for styling, and
JavaScript for functionality, this app allows users to create, edit, delete, and manage notes
efficiently. In addition to text-based notes, the app supports file uploads, enabling users to
attach documents or images to their notes. Each note is also timestamped, displaying both the
creation and modification dates to keep track of changes. The app’s simplicity, combined with
robust features like file management and date tracking, makes it ideal for a wide range of users
seeking an effective note organization tool.
Steps Involved in Building the Online Notes App:
1. Planning and Design:
o Identify the core features of the app: creating, editing, deleting notes, file uploads, and
date tracking.
o Design the layout of the app using wireframes or mockups, ensuring the interface is
intuitive and user-friendly.
o Plan how each element will be structured in HTML, styled with CSS, and interact with
JavaScript for dynamic functionality.
2. Setting Up the HTML Structure:
o Create the basic structure of the app using HTML, including sections like the header,
main note-taking area, and footer.
o Set up input fields for note creation, an area to display existing notes, and buttons for
actions like edit, delete, and file upload.
o Include placeholders for where timestamps will be displayed (creation and modification
dates).
3. Styling the App with CSS:
o Design the layout to be responsive, ensuring the app works well on various screen sizes
(desktop, tablet, mobile).
o Style the buttons, input fields, and note list to ensure the app is visually appealing and
easy to navigate.
o Customize the appearance of notes, highlighting important actions like edit and delete,
and ensuring proper spacing for readability.
4. Adding Interactivity with JavaScript:
o Implement the core functionalities using JavaScript, starting with the ability to add, edit,
and delete notes.
 o Create a system for managing notes dynamically, ensuring that user actions are
reflected immediately in the interface.
o Add file upload functionality to allow users to attach documents and images to their
notes, and ensure these uploads are handled securely and efficiently.
5. Implementing Date Tracking and Testing:
o Use JavaScript to automatically track and display the date and time when notes are
created and modified.
o Test the app thoroughly to ensure all features work correctly, including the note
management system, file uploads, and date tracking.
o Perform usability testing to refine the app’s interface and user experience, ensuring it is
ready for deployment.

Top of Form
Bottom of Form

1.3Project Motivation
The motivation behind developing an online notes app stems from the growing need for efficient,
accessible, and organized note management in both personal and professional environments. In today’s
fast-paced digital world, individuals are constantly juggling tasks, ideas, and important information.
Traditional paper-based note-taking methods or disorganized digital notes often lead to clutter, lost
information, and inefficiency. This project aims to address these challenges by providing users with a
convenient, easy-to-use platform that not only helps them take notes but also keeps everything
organized and accessible in one place.

Another driving force behind this project is the demand for simplicity and customization in note-taking
applications. Many existing tools are either too complex, requiring steep learning curves, or too basic,
lacking the essential features like editing, file attachment, and tracking changes over time. By
incorporating these features into a lightweight and user-friendly app, the goal is to create a solution that
fits into daily workflows seamlessly, without unnecessary complications.

Furthermore, the ability to attach files such as images, documents, and PDFs to notes adds another layer
of versatility to the app. This feature is designed to solve the problem of managing supplementary
materials alongside written notes. For instance, students may want to attach lecture slides to their class
notes, or professionals may need to include supporting documents for meeting notes. By providing this
functionality, the app becomes a comprehensive tool for storing both text-based and multimedia content
in one place.

Tracking the creation and modification dates of notes is another motivating factor. Many users need to
know when a particular note was created or last updated, especially in a professional context where
time-sensitive information is critical. Incorporating automatic date tracking into the app ensures users
always have access to a history of their notes, aiding them in better managing their time and tasks.

Finally, the motivation for choosing HTML, CSS, and JavaScript lies in their accessibility and
flexibility. These technologies are widely supported across platforms, ensuring that the app can be
accessed from any modern web browser without the need for additional installations. Using these tools
also allows the project to be scalable and open for further improvements, making it a versatile solution
that can grow alongside its users’ evolving needs.
Goals and Objectives of the Online Notes App:
Goals:

1. Streamline Note-Taking: The primary goal of the app is to simplify the process of taking,
organizing, and managing notes. By providing an intuitive and user-friendly platform, the app
ensures that users can create, modify, and delete notes quickly and efficiently, without any
unnecessary complexity.
2. Enhance Productivity: The app is designed to boost user productivity by providing an
organized and centralized location for all notes. Whether for personal or professional use, users
can easily access, update, and manage their notes, helping them stay on top of tasks and
information.
3. Provide Versatility with File Uploads: By allowing users to upload and attach files (such as
documents and images) to their notes, the app expands its functionality beyond text-only notes,
making it a comprehensive tool for managing various types of information.
4. Enable Effective Time Management: The app includes date and time tracking for note
creation and modification, helping users manage their tasks more effectively by keeping a clear
history of when notes were added or updated.
5. Ensure Cross-Platform Accessibility: As a web-based application, the app is designed to be
accessible across multiple devices and platforms without the need for installation. This ensures
that users can access their notes from anywhere, at any time, through a web browser.

Objectives:
1. User-Friendly Interface: Develop an easy-to-navigate interface using HTML and CSS,
ensuring that users can interact with the app seamlessly. The layout should be simple, with clear
input fields and buttons for adding, editing, deleting, and managing notes.
2. Efficient Note Management: Implement the ability to add, edit, and delete notes using
JavaScript, ensuring that these actions are carried out smoothly and efficiently. Notes should
update in real-time, offering a seamless user experience.
3. File Attachment Support: Provide a file upload feature that allows users to attach relevant files
(documents, images) to specific notes, broadening the functionality of the app and making it
more versatile for different use cases.
4. Track Creation and Modification Dates: Include a date and time feature that automatically
tracks when a note is created or updated. This ensures that users can keep an accurate record of
their notes and updates.
5. Cross-Device Compatibility: Ensure the app is fully responsive and compatible with different
screen sizes and devices, such as desktops, tablets, and mobile phones, to make it accessible
from anywhere at any time.
 LISTOFFIGURES

FIGURE FIGURESNAME PAGENO

S .
1.1 StructureOfBloodCells 2
1.2 RedBloodCells 7
1.3 WhiteBloodCells 7
2.1 Fold_3Dataset 12
2.2 Fold_3Precision 13
2.3 Dataset images 15
4.1 DatasetVersion 26
5.1 Indentation BlockDiagram 28
5.2 WorkingOf Graphics 30
ProcessingUnit
5.3 Epoches Compression 34

LISTOFTABLES
TABLES TABLENAME PAGENO
.
DifferenceBetweenRedBloodcellsAndWhiteB
1.1 lood Cells 2
6.1 ConfusionMatrix 38
6.2 Quantitative Comparision 40
1. INTRODUCTION

Blood is a specific type of circulating connective fluid that takes oxygen from the lungs
andtransportsittoallhumanbodycells.Thebody'scellsrequireoxygenformetabolism,whichthebloodcarriesfrom
the lungs to the cells. In this way, blood nourishes cells carries hormones, and eliminates
unwantedmaterialsthatareeventuallyeliminatedbyorgansliketheliver,kidneys,orintestine.Additionally,theblo
odreturns the carbon dioxide created during metabolism to the lungs, where it is expelled out. The blood
iscomprisedofplasmathatformsliquidportionandcellfragments.Thecellfragmentsarecomposedofwhitebloodc
ells(wbcs)with1%proportion,responsibleforimmunity.Redbloodcells(rbcs)makeup40-50%oftotal blood
volume, delivering oxygen and carbon dioxide, and platelets are responsible for blood
clotting .Wbcscanbedividedintotwogeneral groupsbasedonthepresenceof
granules:granulocytesandagranulocytes(non-
granulocytes).Lymphocytesandmonocytesfallunderagranulocytes,whileneutrophils,eosinophils,andbasophil
sareconsideredtobegranulocytes.Undevelopedwbcscalledimmaturegranulocytes (IG) are expelled from the
bone marrow into the blood. The presence of IG (promyelocytes,myelocytes, and metamyelocytes) in the
blood signifies an early reaction to an infection, swelling, or someother type of problem with the bone
marrow like leukemia, except the blood from newly born children orexpectingwomen.
Whitebloodcellsare
infectionfighters;todestroyforeignproteinsfoundinbacteria,viruses,andfungi,wbcsdividethemselvestofight
againstinfectionsanddiseasesbydetecting,recognizing,andbindingthemselvestoforeignproteins[3].Redbloodc
ells,alsoknownaserythroblasts,helptissuesproduceenergyby delivering appropriate oxygen. When energy is
produced, waste in the form of carbon dioxide is
alsoformed.Rbcsareresponsibleforprovidingthatcarbondioxidetothelungssothatitisexhaled.Erythroblastsare
immature rbcs that are usually present in the blood of newly born child in a duration from 0-4 months.Their
presence in human blood after the neonatal period (0-4 months) indicates severe problems
likedamagedbonemarrow,stress,andmalignanttumourthatmayleadtocancerorbenignthatgrowsinsizebutdo not
infect other body parts. Platelets are also known as thrombocytes and are essential for the
immunitysystem;theirprimaryresponsibilityistostopbleeding.
If bleeding starts from an injury or blood vessel damage somewhere in the body, the brain
sendsan alerting signal to the platelets. The platelets flow to the wounded area, cluster together, and form a
clot,sealingthebloodvesseltostopthebleeding.Theyalsoplayanimportantroleintissuerepairandremodellingtopr
eventtumourprogressionandleakageofvesicantfluids.Theycompriseatinyproportion,
i.e.,lessthan1%ofbloodvolume.Typically,thecommonratesofneutrophilsinthebloodare0-
6%.Eosinophilscomprise1-3%, basophils 0-1%, lymphocytes 25-33%, and monocytes 3-10% of the
leukocytes floating in the blood[4]. The classification of blood cells is a current research area for scientists
trying to diagnose diseases
thataffectbloodcells.Bloodcellclassificationusingmicroscopicimagesofbloodwouldonlybedonemanuallybym
edicalprofessionalswiththe necessaryexperienceandtraining.
Blood is analysed in two different ways. The first method is a complete blood count (CBS)
testthat calculates the total percentage of rbcs, WBC, and platelets; the second is the peripheral blood
smears(PBS) test. These results represent the patient's overall health. The rbcs, WBC, and platelets types
and anearly diagnosis of the diseases can be determined using microscopic blood images. Each sort of cell
in thehumanbloodhasapurpose.Achangeinthenumber
ofbloodcelltypeswouldresultinanillnessordisease.Manyillnesses,includingbloodcancer,canbe
broughtonbyalowWBCcount.
A lower count of healthy red blood cells leads to Anemia and a lower ratio of platelets leads
toexcessive bleeding and bruising. Conventional blood cell type detection methods take a long time and
havelow accuracy, which highlights the significance of accurate systems for the rapid and precise analysis
ofblood cells . Blood cells in microscopic images of blood smears have been categorized using
conventionalmachine learning (ML) techniques like support vector machine, decision tree, k-nearest
neighbor, naiveBayes, and artificial neural network. The general process flow for traditional ML
approaches includes pre-
processingofbloodsmearimages,segmentationtodividethecell,featureextraction,featureselectionto
remove undesired data, and classification. Despite many promising results, feature extraction and
selectionsignificantlyaffecthow wellclassicalMLalgorithmsperforminclassification.
Choosingtheoptimalfeaturesandfindingtheappropriatefeatureextractionalgorithmhasbecomecompl
ex and time-consuming . Several deep learning (DL) techniques of convolutional neural networks(cnns)
have recently been proposed to tackle this challenging topic. Recent developments in deep learningallow us
to estimate the type of blood cells from microscopic images. In contrast to conventional MLapproaches,
DL-based approaches have the capability of autonomous feature extraction and selection.
Priorresearchrevealedthatfor classifyingbloodcells,CNNperformedbetterthantraditionalMLtechniques.

Fig: -1.1StructureofBloodcells

StructureandFunctionofBlood

● RedBloodCells(RBCs):Biconcave,anucleate,filledwithhemoglobin.
● WhiteBloodCells(WBCs):

● Granulocytes: Neutrophils (multi-lobed nucleus, fine granules), Eosinophils (bi-

lobednucleus,largered-orangegranules),Basophils(bi-lobednucleus,largeblue-
purplegranules).
● Agranulocytes:Lymphocytes(largenucleus,thincytoplasm),Monocytes(kidney-
shapednucleus,abundantcytoplasm).

● Platelets(Thrombocytes):Small,disc-
shapedcellfragmentsderivedfrommegakaryocytes,containinggranulesforclotting.

Eachtypeofbloodcellisuniquelystructuredtoperformitsspecificfunctions,contributingtotheoverallhealthand
homeostasis ofthebody.

ProjectOverview:

Objective:Theprimary
goalofthisprojectistodevelopanautomatedsystemtodetectandclassifyRedBloodCells(RBCs)andWhiteBloodCe
lls(WBCs)inmicroscopicbloodsmearimagesusingdeeplearning
techniques.Thissystemaimstoassistinthediagnosisandmonitoringofvariousblood-
relateddiseasessuchasanaemia,infections,andleukaemia.
StepsInvolved:

1. DataCollection:

Acquire a dataset of labelled blood smear images containing RBCs and WBCs. Public
medicalimage databasesordatasets providedbyhealthcareinstitutionscanbe used.

2. DataPreprocessing:

Preprocess the images to enhance quality and consistency. This includes resizing,
normalization,and augmentation to increase the dataset size and variability, which helps in improving the
model'srobustness.

3. ModelSelection:

Chooseanappropriatedeeplearningarchitectureforimageclassificationtasks,suchasConvolutionalN
euralNetworks(CNNs).CommonlyusedmodelsincludeResNet,VGG,orcustomCNNarchitectures.

4. ModelTraining:

Train the selected deep learning model on the pre-processed dataset. Use a portion of the data
fortrainingandanotherportionforvalidationtomonitorthemodel'sperformanceandpreventoverfitting.

5. Evaluation:

Evaluatethemodel'sperformanceusingmetricssuchasaccuracy,precision, recall,andF1-score.
EnsurethemodelcanaccuratelydetectandclassifyRBCsandWBCs.

6. Fine-tuningandOptimization:

Fine-tune the model by adjusting hyperparameters, using techniques like transfer learning,
oraddingadditionallayerstoimproveaccuracy.Performcross-validationtoensurethemodel'sgeneralizability.

7. Deployment:

Deploy the trained model into a user-friendly application or system that can process new
bloodsmear images in real-time. This could be a standalone software tool or integrated into existing
medicaldiagnostic systems.

8. Validation and Testing:

Conductextensivetestingandvalidationwithnew,unseendatatoensurethemodel'sreliabilityinaclinica
lsetting.Obtain feedbackfrommedicalprofessionalsanditerativelyimprovethesystem.
 Applications:

● ClinicalDiagnostics:Assistspathologistsindiagnosing blooddisordersmoreefficiently.
● Research: Providesatoolfor researcherstoanalyzelargedatasetsofbloodsamples.
● Education:Canbeusedasaneducationaltoolformedicalstudentslearning hematology.

Challenges:

● DataQualityandAvailability:High-
qualitylabeleddatasetsareessentialbutcanbechallengingtoobtain.
● ModelAccuracy:Achievinghighaccuracyiscritical,asmisclassificationcanleadtoincorrectdiagnose
s.
● DeploymentinClinicalSettings:Ensuringthemodel'srobustnessandreliabilityinreal-
worldclinicalenvironmentsiscrucial.

By automating the detection and classification of RBCs and WBCs, this project has the potential
tosignificantly enhance the efficiency and accuracy ofhaematological analysis, benefiting both
medicalprofessionalsandpatients.

ProjectMotivation:

TheprimarymotivationbehinddevelopingasystemforthedetectionofRedBloodCells(RBCs)andWhiteBloodC
ells(WBCs)usingdeeplearningtechniquesstemsfromseveralkeyfactors:

1. MedicalImportance

●DiagnosticAccuracy:AccuratedetectionandcountingofRBCsandWBCsarecrucialfordiagnosingvariou
smedicalconditionssuchasanemia,infections,andleukemias.Manualcountingunderamicroscopeisprone
tohumanerrorandvariability,whichcanaffectdiagnostic accuracy.
● DiseaseMonitoring:RegularmonitoringofRBCandWBCcountsisessentialforpatientsundergoingtreatm
ent for diseases like cancer. Automated systems can provide consistent and reliable results,
ensuringbettermonitoringandmanagementofsuchconditions.

2. EfficiencyandSpeed

●Time-Consuming Manual Process: Traditional methods of cell counting are labor-intensive

andtime-consuming. Automating this process with deep learning can significantly reduce the time required
foranalysis,allowingforfasterdiagnosisandtreatment.
● HighThroughput:Automateddetectionsystemscanprocesslargevolumesofbloodsamplesquickly,whic
hisparticularlybeneficialinclinicalsettingswithhighpatientloads.

3. TechnologicalAdvancements

Deep Learning Capabilities: Advances in deep learning have shown remarkable success in
●
imagerecognitiontasks.Utilizingthesetechniquesformedicalimageanalysiscanleveragetheircapabilitytolearnc
omplexpatterns andfeatures frombloodsmearimages.
 ● Improved Accuracy: Deep learning models can achieve high accuracy in detecting and
classifyingcells by learning from large datasets of labeled images, reducing the error rates associated with
manualmethods.
 4. Cost-Effectiveness

● Reducing Costs: Automated systems can lower the costs associated with manual labor and the
needfor multiple retests due to human error. This can make diagnostic services more affordable and
accessible,especiallyinresource-limitedsettings.

5. EnhancedHealthcareDelivery

● Standardization:Automateddetectionensuresstandardizedanalysisacrossdifferentlaboratoriesandtech
nicians,leadingtomore consistentandreliablehealthcare delivery.
● RemoteDiagnostics:Suchsystemscanbeintegratedwithtelemedicineplatforms,allowingforremotediagn
osticsinareaswhereaccesstohealthcare professionalsislimited.

By addressing these motivations, a deep learning-based system for RBC and WBC detection aims
toenhancetheaccuracy,efficiency,andaccessibilityofbloodcellanalysis,ultimatelycontributingtoimprovedpati
entoutcomes andhealthcaredelivery.

Goals&objectives

Objective

Theprimarygoalofthisprojectistodevelopadeep learning modelthat canaccuratelydetectandclassifyRed

Blood Cells (RBCs) and White Blood Cells (WBCs) in microscopic blood images. This can assistmedical
professionals in diagnosing and monitoring various health conditions, such as anemia,
infections,andblooddisorders.

SpecificGoals:

DataCollectionandPreprocessing:

GatheradiversedatasetofmicroscopicbloodimagescontaininglabeledRBCsandWBCs.

Preprocesstheimagestoenhancequalityandstandardizethe
inputforthedeeplearningmodel,includingresizing,normalization,andaugmentation.

ModelDevelopment:

Designandimplementaconvolutionalneuralnetwork(CNN)architecturesuitableforimageclassificationanddet
ectiontasks.

Experimentwithvariousarchitectures(e.g.,VGG,ResNet,U-Net)toidentifythemosteffectivemodelforthetask.

TrainingandValidation:

Trainthemodelusingthepreprocesseddataset,ensuringapropersplitbetweentraining,validation,andtestsets.

Utilizetechniquessuchascross-validation,earlystopping,anddataaugmentationtoimprovemodelperformance
andpreventoverfitting.
 1. EvaluationandOptimization:

Evaluatethemodel'sperformanceusingmetricssuchasaccuracy,precision,recall,F1-
score,andconfusionmatrix.Optimizethemodelbytuninghyperparameters,adjustingthearchitecture,andempl
oyingtechniqueslike transferlearningifnecessary.

2. DetectionandClassification:

Implement the trained model to detect and classify RBCs and WBCs in new microscopic blood
images.Ensure the model can differentiate between different types of WBCs (e.g., neutrophils,
lymphocytes,monocytes)ifrequiredbythe projectscope.

3. Deployment:

Developauser-friendlyinterfaceorapplicationtodeploy themodelforpractical usebymedicalprofessionals.

Ensure the application provides visualizations and detailed reports of the detection
andclassificationresults.

4. TestingandValidationinReal-worldScenarios:

Test the deployed model in real-world clinical settings to validate its effectiveness and reliability.
Gatherfeedback from medical professionals to further refine and improve the system. By achieving these
goals,the project aims to provide a reliable and efficient tool for automated RBC and WBC detection,
aiding inthetimelyandaccurate diagnosis ofvariousmedicalconditions.

DifferenceBetweenRedBloodCellsandWhiteBloodCells

Both red blood cells and white blood cells play an essential role in the human body. Red blood cells
orRBC carry oxygen to the tissues in different parts of the body. White blood cells or WBC strengthen
thedefense mechanism of the body by generating antibodies. The primary difference between RBC and
WBCliesintheirfunctionality.WhileRBCsactascarriers,WBCactascreators.Theexpertsaresummarizedthediffe
rencebetweenredbloodcellsandwhite blood cellsina comprehensivemanner.

RedBlood Cells(RBC)
As the name suggests, RBC is red because of the presence of haemoglobin which is an iron-rich
proteinand binds with oxygen to get the red colour. RBC gives a red colour to the blood because of its
presence inthe blood in a large number. Also known as Erythrocytes, red blood cells are round, small, and
bi-concavedinshapebutduetotheir flexibility, theyappearbell-shapedwhenpassingthroughsmallvessels.
Theycarryoxygen to the tissue in the body. To maintain a healthy RBC count in the body, it is essential to
take an ironand vitamin-rich diet. A low RBC count causes anaemia and its common symptoms are
irregular heartbeat,pale skin, feeling cold, fatigue, and joint pain. The primary function of Red Blood Cells
is to carry
oxygenfromthelungstothetissueindifferentpartsofthebody,usingthebloodcirculationsystem.Theyalsocarrycar
bon dioxide back to the lungs from where they are excreted out of the body. Since the RBC has a bi-
concaveshapeithelpsinthe exchange ofoxygenataconstantrateandovera large surfacearea.
 Fig:-1.2RedBloodCells
WhiteBloodCells(WBC)
WhiteBloodCellsarecolourlessduetotheabsenceofhaemoglobininthem.AlsoknownasLeukocytes,white
blood cells protect the body from any infections by producing antibodies that build up the definessystem of
thebody againstgerms andinfections. One of the otherimportantfactors that helpus
todifferentiatebetweenRBCandWBCisthecirculationsystemusedbythesecells.
WBCusescardiovascularcirculation andis alsopresentin thelymphaticsystem. Red blood cells use only the
cardiovascularcirculatory system. Invading bacteria, viruses, and germs are attacked by these cells, which
aid in the fightagainstinfection.

Fig:1.3WhiteBlood cells
Although white blood cells begin in the bone marrow, they circulate throughout the body. There are
fivedifferenttypes ofwhitebloodcells:

● Neutrophils

● Lymphocytes

● Eosinophils

● Monocytes

● Basophils
 Theprimaryfunctionofwhitebloodcellsisto produceantibodiesinthebodyandto
strengthenimmunityofthebody.Agooddefinesmechanismprotectsthebodyfromanygermattacksorinfections.W
BCprotectsthebodybydigesting theforeignmaterialand cancercellspresentinthe bodybyproducingantibodies.

Let’s lookatthedifferencebetweenredbloodcellsandwhitebloodcells indetail.

DifferenceBetweenRedBloodCellsandWhiteBloodCells

Criteria RBC WBC

RBC is scientifically
Scientificname WBC iscalledLeukocytes.
calledErythrocytes.

RBCs are a
Appearance nucleated,bi- WBCarenucleatedandirregularinshape.
concave,and
disc-shaped.

ThesizeofRBCisroughly 6-
Size ThesizeofWBCis15microns.
8microns.

Producti RBCisproducedintheredbonema WBCisproducedinthespleen,lymphnodes,

onlocation rrow. etc.

Producti Almost 2 million RBC WBCareproducedinacomparativelylowern

onnumber areproduced inthe umberthanthe RBC.
bodypersecond.

Formati TheprocessofRBCformationisca TheprocessofWBCformationin

onprocess lledErythropoiesis. thebodyiscalledleucopoiesis.

Motility RedBloodCellsarenon-motile. Whitebloodcellsare motile

RBCaccountsfor36%-
50%oftheblood in the body. This
Percentage Incomparison,WBCconstitutesameagre1
percentage,however,differsaccord
%oftheblood.
ingtotheheight,weight,andageofth
inblood
eperson.
 White blood cells are of multiple types.
T-lymphocytes, B-lymphocytes (plasma
Redbloodcellsareonlyofonetype.
Types cells),monocytes(macrophages),neutrophils
,eosinophils,and basophils are some of
thetypesofWBC.

RBCcansurviveupto120daysinth WBC can survive anywhere

Lifespan
e body.
betweenseveraldaystoevenseveralyearsinth
ebody.

RBC is made up of WBCismadeupofantibodieswithMHCanti

Constitution
onlyhaemoglobin. gencellmarkers.

Thepresenceofhaemoglobinlend
TheabsenceofhaemoglobinmakesWBCco
s a red colour to the RBC. It isthe
Colour lourless.
reason they are called red
bloodcells.

The primary function of the TheprimaryfunctionofWBCistoproducea

RBCistocarryoxygentothevarious ntibodiestostrengthenthedefinesmechanism
parts of the body. As a ofthebody.Theseantibodiesprotectthebodyfr
Function
secondaryfunction,theyalsocarryw omanyattackbygermsandprovide immunity
astematerials and carbon dioxide against infections. Some
to thelungs. ofthemarealsophagocytic.

The circulation system used is

The circulation systems used
Circulation thecardiovascularsystemthatisrelate
dtothe bloodvesselsandthe heart.
arecardiovascularaswellaslymphatic.
 A low RBC count in the body
A low WBC count can lead to
canlead to anaemia which can
Lowcounteffec leukopeniathat can hamper the immune
affect
t system of thebody.
thebody’sabilitytocarryandsupplyox
ygentothetissues.

A high WBC countis an indication

AhighRBCcountisproducedintheb
ofinfection present in the body or of a
Highcounteffect odyduringexerciseorathighaltitudes.
lowerresponse rate of the bone marrow.
Such acondition iscalledLeucocytosis.
 2.DATASET
hedatasetforRedBloodCell(RBC)andWhiteBloodCell(WBC)detectionusingdeeplearningtypicallyconsists
of microscopic images of blood samples. Here's a brief overview of what such a dataset mightinclude:

ComponentsoftheDataset

1. MicroscopicImages:

● High-resolutionimagescapturedusingamicroscope.

● Imagescanbestainedwithvariousdyestohighlightdifferentcomponentsofthebloodcells.

2. Annotations:

● BoundingBoxes:CoordinatesthatspecifythelocationofRBCsandWBCswithineachimage.

● Labels:Classlabelsindicatingwhetherthedetectedcellis anRBC,WBC, orothertypesofcells.

● Segmentation Masks:Pixel-
wiseannotationsthatpreciselyoutlinetheboundariesofthecells(incaseofsegmentationtasks).

3. Metadata:

● Informationaboutthesample, suchasthepatient’s age,gender, medicalcondition, etc.

● Detailsaboutthestainingmethodand imagingconditions.

Characteristics

● Diversity:Imagesshouldrepresentadiverserangeofconditionsandvariationstoensurethemodelgener
alizes well.
● ClassBalance:ThedatasetshouldideallyhaveabalancednumberofRBCs
andWBCstoavoidbiastowardsonetypeofcell.
● Quality:High-quality,clearimagesareessentialforaccuratedetectionandclassification.

UsageinDeepLearning

● Training:Theannotatedimagesareusedtotraindeeplearningmodels,suchasconvolutionalneuralnetw
orks(CNNs),todetectandclassifyRBCsandWBCs.
● ValidationandTesting:Separatesetsofimagesareusedtovalidateandtestthemodel’sperformance,en
suringitcangeneralize tonew,unseendata.

CommonDatasets

● BCCDDataset:Acommonlyuseddatasetforbloodcelldetection,whichincludesimageswithannotatio
 nsforRBCs,WBCs,andplatelets.
● ISBIChallengeDatasets:Datasetsprovidedforspecificchallengesandcompetitionsfocusingonbiom
edicalimageanalysis.
 Applications

● MedicalDiagnosis:AutomateddetectionandcountingofRBCsandWBCscanassistindiagnosingvari
ousblood-relatedconditions,suchasanemia,infections,andleukemia.
● Research:Helpsinstudyingthecharacteristicsandbehaviorofbloodcellsunderdifferentconditions.

Using deep learning for RBC and WBC detection enhances the accuracy and efficiency of
medicaldiagnostics,reducing the workload on pathologists and enablingfasterdecision-makingin
clinicalsettings.

DatasetUsedforDetectionofRedBloodCellsandWhiteBlood

Fig:2.1Fold_3Dataset

The "Fold 3 by V8*V9 Dataset of RBC and WBC Detection Using Deep Learning" likely refers to
aspecific configuration or fold of a dataset used in a cross-validation process for training and
evaluatingdeep learning models to detect Red Blood Cells (RBCs) and White Blood Cells (WBCs).
Here’s a briefexplanationofthekeyelementsinvolved:

1. Dataset:

● RBCandWBCDetection:Thedatasetcontainsimagesofbloodsamples,annotatedwithlabels
identifying RBCs and WBCs. This data is used to train models to
automaticallydetectandclassifythese cells.
 ● V8*V9: This could denote a specific version or preprocessing pipeline applied to
thedataset.Eachversion(V8,V9,etc.)mightincludedifferentaugmentations,resolutions,orpre
processingtechniques.

2. Fold3:

● Cross-Validation: Cross-validation is a technique used to assess the performance

andgeneralizability of a model. The dataset is split into multiple folds (e.g., 5 or 10). In
eachiteration, one fold is used for validation/testing, while the remaining folds are used
fortraining.
● Fold 3: Indicates that the third segment of the dataset is used as the validation set,
whiletherestareusedfortraininginthisparticularrun.Thisprocessisrepeatedforeachfoldtoensu
re themodelis testedonallparts ofthedataset.

3. DeepLearning:

● Model Architecture: Various deep learning models (such as CNNs) are used for
imagedetection and classification tasks. These models are trained to learn the features of
RBCsandWBCsfromtheimages.
● TrainingandEvaluation:Themodelistrainedonthetraining
foldsandevaluatedonthevalidation fold (Fold 3 in this case). Performance metrics such as
accuracy, precision,recall, and F1-score are calculated to measure the model’s
effectiveness in detecting andclassifyingthecells.

Overall, the "Fold 3 by V8*V9 Dataset" approach helps in creating robust models that are well-
validatedandless pronetooverfitting,ensuringtheyperformwellonunseendata.

Fig:2.2Fold_3Precision
DatasetNameAndDetails

Name :-
RBC&WBCFold_3
Modal :-YoloV8
Totalclass :-02
TotalImages :-1000
FrameworkTool :-Robflow
TrainImages :-960
ValidImages :-0
TestImages :-0

1. DataPreprocessing:

● Imagenormalizationandaugmentationtoenhancethedatasetandimprovethemodel'srobustnes
s.
● AnnotationsorlabelsforRBCsand WBCstosupervisethe learningprocess.

2. ModelTraining:

● UsingadeeplearningframeworklikeTensorFloworPyTorchtobuildandtrainthemodel.

● Implementingtechniqueslikedropoutandbatch normalization
topreventoverfittingandimprove generalization.

3. ModelEvaluation:

● Usingcross-validation(e.g., Fold3)toevaluatethemodel'sperformance.

● Analyzingtheconfusionmatrixtounderstandthemodel'sperformanceindifferentclasses.

ThisbriefoverviewprovidesageneralunderstandingoftheFold3byV8*V9DatasetforRBCandWBCDetectio
nusingDeepLearning.Specificdetailsmayvarybasedontheactualdatasetandthedeeplearningtechniquesappli
ed.
Fig:-2.3DataProcessingflowchart
 Fig:-2.4Datasetimages
DatasetSource:

RoboFlowisaplatformspecificallydesigned forcomputervisionandmachinelearningprojects.Itoffersa
variety of datasets, tools, and resources to help you build and deploy computer vision models.
Here'showyoucanleverage RoboFlowforyourprojects:

RoboFlowDatasets

RoboFlowprovidesaccesstonumerouspublicdatasets,andyoucanalsouploadyourowndatasets
forannotationandpreprocessing.

● RoboFlowUniverse:Acommunity-
drivencollectionofpublicdatasetsforvariouscomputervisiontaskslikeobjectdetection,classification,
andsegmentation.

● RoboFlowUniverse

KeyFeatures
1. DatasetManagement:Upload,organize,andmanageyourdatasetswithease.RoboFlowsupportsvari
ousfileformatsandprovidestoolsfor datasetaugmentationandpreprocessing.
2. Annotation Tools:Usebuilt-in annotation toolstolabel yourimagesfor tasks such as
objectdetection,segmentation,andclassification.
3. AugmentationandPreprocessing:Automaticallyaugmentandpreprocessyourimagestoimprovem
odelperformance.Thisincludesresizing,rotating,flipping,andmore.
4. ModelTraining:IntegratewithpopularmachinelearningframeworkslikeTensorFlow,PyTorch,and
YOLO totrainyourmodelsdirectlyfromtheplatform.
5. Deployment:Deployyourtrainedmodelstoproductionenvironments,includingmobileandwebapplic
ations,usingRoboFlow’sdeploymentoptions.
 HowtoUse RoboFlow

1. Sign UpandCreatea Project: Createafreeaccount onRoboFlow and startanewproject.

2. UploadYourData:ImportyourdatasetorexplorepublicdatasetsavailableonRoboFlowUniverse.
3. AnnotateandPreprocess:Usetheannotationtoolstolabelyourdataandapplyaugmentationtechnique
s.
4. TrainYourModel:UseRoboFlow’sintegrationwithvariousmachinelearningframeworkstotrainyou
rmodel.
5. DeployYourModel:DeployyourmodelusingRoboFlow’sdeploymentsolutions.

DatasetComposition

In a project involving the detection of Red Blood Cells (RBCs) and White Blood Cells (WBCs)
usingdeep learning, the types and quantities of images available are crucial for training, validating, and
testingthe model.Hereisa detailedoutlinebasedoncommondatasets used insuchprojects:

TypesofImages

1. MicroscopicBloodSmearImages:

● Description: High-resolution images of blood smears taken using a microscope.

Theseimagesincludevarioustypesofblood cellssuchasRBCs,WBCs,and platelets.
● Annotation:Typically,theseimagesareannotatedtomarkdifferentcelltypes.Annotationscan
be donemanuallybyexpertsorusingsemi-automatedtools.

2. SegmentedCellImages:

● Description:Theseimagesfocusonindividualcellsextractedfromthelargerbloodsmearimage
s.Theycan beusedto train modelsspecificallyfortheclassificationofcelltypes.
● Annotation: Each segmented cell image is labeled as RBC, WBC (and further
classifiedintosubtypeslikeneutrophils,lymphocytes,monocytes,eosinophils,basophils),orpl
atelets.

QuantitiesofImages

Thequantityofimagesineachcategorycanvarybasedonthedataset.Herearesometypicaldatasetsandtheirquan
tities:

1. BCCD(BloodCellCountandDetection) Dataset:

● TotalImages:Approximately364images.

● RBCs:Containsaround29,863labeledRBCs.

● WBCs:Containsaround3,512labeledWBCs.

● Platelets:Containsaround8,696labeledplatelets.
 2. LISCDataset (LeukocyteImagesforSegmentationandClassification):

● TotalImages:Contains240imagesofbloodsmears.

● WBCs:ContainsannotationsforvarioustypesofWBCs.

3. ALL-IDB(AcuteLymphoblasticLeukemiaImageDatabase):

● TotalImages:Contains260images.

● WBCs:PrimarilyfocusedonWBCs,especiallyleukemiccells.

● RBCsandOtherComponents:Alsoincludesnormalbloodcomponentsforcontext.

SummaryofDatasetQuantities

● TotalNumberofImages:

● BCCD:364images

● LISC:240images

● ALL-IDB:260 images

● AnnotatedBloodCells:

● BCCD:29,863RBCs,3,512WBCs, 8,696Platelets

● LISC:VariousWBCs

● ALL-IDB:FocusedonWBCs(bothnormalandleukemiccells)

SourceAccess

Thesedatasetsarepubliclyavailableandcanbeaccessedfromvarioussources:

● BCCDDataset:AvailableonGitHubandKaggle.
● LISCDataset:Availableonresearchdatabasesanduniversityrepositories.
● ALL-IDBDataset:Availableondedicatedleukemiaresearchdatabasesandacademicpublications.

Formoredetailedinformationandaccesstothesedatasets,youcanrefertothefollowinglinks:

● BCCDDataset:GitHubRepository
 ● LISCDataset:UniversityRepository
● ALL-IDBDataset:ALL-IDBDatabase

Byusingthesedatasets,youcantraindeeplearning modelseffectivelyforthedetectionandclassificationof
RBCs and WBCs, leveraging the annotated images to improve the accuracy and reliability of
yourmodels.

DataPreprocessing
Preprocessing the dataset for a project focused on detecting Red Blood Cells (RBCs) and White
BloodCells (WBCs) using deep learning is a crucial step that involves several stages. This process aims
toenhance the quality and usability of the data, ultimately improving the performance of the deep
learningmodel.Hereare thekeystepsinvolvedinpreprocessingsuchadataset.
1. DataCollection

● Source: Obtain a dataset from a reliable source such as medical research databases,
publicrepositories (e.g., Kaggle, UCI Machine Learning Repository), or collaborations with
medicalinstitutions.
● Composition: Ensure the dataset includes labeled images of blood smears with annotations
forRBCsandWBCs.

2. DataCleaning

● QualityCheck:Reviewimagesforquality,removinganythatareblurry,corrupted,orirrelevant.
● LabelVerification:Verifythatallimagesarecorrectlylabeledandannotated.

3. DataAnnotation

● ManualAnnotation:Ifthedatasetisnotpre-annotated,useannotationtools(e.g.,Labelbox,VGG
Image Annotator)tomanuallylabelRBCsandWBCsineachimage.
● AutomatedAnnotation:Usepre-
trainedmodelstoassistwithinitialannotations,whichcanthenbemanuallyrefined.

4. DataAugmentation

● Purpose:Augmentationincreasesthediversityofthedataset,helpingthemodelgeneralizebetter.
● Techniques:

Rotation:Randomlyrotateimagesbysmalldegrees.

Flipping:Applyhorizontalandverticalflips.

Scaling: Scaleimagesupordown.

Translation: Shiftimagesalongthexoryaxis.

5. Normalization

● PixelScaling:Normalizepixelvaluestoarangeof[0,1]or[-
1,1]tofacilitatefasterandmoreefficienttraining.
● MeanSubtraction:Subtractthemeanpixelvalueanddividebythestandarddeviationifusingcertaindee
plearningframeworks.

6. ImageResizing

● ConsistentSize:Resizeallimagestoaconsistentsize(e.g.,128x128,256x256)toensureuniformityinth
einputdimensionsfortheneuralnetwork.
7. DataSplitting

● TrainingSet: Typically70-80%ofthe dataset.

● ValidationSet: Typically10-15%ofthe dataset.
● TestSet:Typically10-15%ofthedataset.
● StratifiedSplit:Ensurethatthesplitmaintainstheproportionofdifferentclasses(RBCsandWBCs).

8. DataBalancing

● Class Imbalance: Address any class imbalance issues by oversampling the minority class,
undersamplingthemajorityclass,orusingtechniqueslikeSMOTE(SyntheticMinorityOver-
samplingTechnique).
 3. LiteratureSurvey

Introduction

The detection and classification of Red Blood Cells (RBCs) and White Blood Cells (WBCs) are
criticaltasksinmedicaldiagnostics.Traditionalmethodsarelabor-
intensiveandpronetohumanerror,necessitatingthedevelopmentofautomatedsystemsusingdeeplearningtech
niques.Thisliteraturesurveyexplores various approaches and advancements in the application of deep
learning for detecting andclassifyingRBCs andWBCs.

TraditionalApproaches

Traditional methods for blood cell detection involve manual counting and visual inspection under
amicroscope,whicharetime-
consumingandsubjecttovariabilitybetweenobservers.Automatedmethodsusing image processing
techniques have been developed to address these issues, but they often lack therobustness
andaccuracyrequiredforclinicaluse.

DeepLearninginMedicalImaging

Deeplearning,asubsetofmachinelearning,hasshowngreatpromiseinmedicalimagingduetoitsabilityto
automatically learn features from raw data. Convolutional Neural Networks (CNNs) are
particularlyeffectiveforimageclassificationandsegmentationtasks,
makingthemsuitableforbloodcelldetection.

KeyStudiesandApproaches

1. CNN-basedDetectionandClassification

● Zhou et al. (2018) developed a CNN-based method for detecting and classifying
RBCsand WBCs. Their approach involved using a deep CNN to extract features from
bloodsmear images and then classifying the cells using a softmax layer. The model
achievedhigh accuracy and demonstrated the potential of deep learning in automating
blood cellanalysis

2. U-NetArchitectureforSegmentation

● Ronnebergeret al. (2015)introducedtheU-Netarchitecture, whichhasbeenwidelyusedfor

medical image segmentation. This architecture has been applied to segment WBCsfrom
complex backgrounds in microscopic images, achieving remarkable accuracy
androbustnessindiversedatasets

3. FeatureFusionTechniques

● Huang et al. (2020) proposed a novel feature fusion-based deep learning framework
forWBCclassification.ThismethodcombinedfeaturesfrommultipleCNNlayerstoenhancethe
classification performance, particularly in distinguishing between different types
ofWBCs
 4. TransferLearning

● Nguyen et al. (2019) demonstrated the effectiveness of transfer learning for blood
cellclassification.Byfine-tuningpre-
trainedCNNmodelsonbloodcelldatasets,theyachievedhighclassificationaccuracywithreduc
edtrainingtime,highlightingtheefficiencyoftransferlearninginmedicalimagingapplications.

5. AttentionMechanisms

● Jiang et al. (2018) incorporated attention mechanisms into their CNN model to
improveWBC classification. The attention module allowed the model to focus on
relevant parts ofthe image,therebyenhancingthe accuracyofcelldetectionandclassification

EvaluationMetrics
The performance of deep learning models for blood cell detection is typically evaluated using
metricssuchasaccuracy,precision,recall,F1-
score,andconfusionmatrices.Thesemetricsprovideacomprehensiveassessmentofthemodel'scapabilitytocor
rectlyidentifyand classify bloodcells.

ChallengesandFutureDirections

Despite significant advancements, challenges remain in the development of deep learning models
forbloodcelldetection.Theseinclude:

● DatasetDiversity:Theneedfordiverseandrepresentativedatasetstotrainrobustmodels.
● Generalization:Ensuringmodelsgeneralizewelltodifferentstainingtechniquesandimagingconditio
ns.
● Real-timeProcessing:Developingefficientmodelscapableofreal-
timeanalysisforclinicalapplications.

Future research directions include exploring advanced architectures such as Transformer

networks,improvingdataaugmentationtechniques,andintegratingmulti-
modaldatatoenhancemodelperformance.

Conclusion

The application of deep learning to the detection and classification of RBCs and WBCs holds
greatpotential for improving the accuracy and efficiency of medical diagnostics. Continued research
anddevelopment in this field are likely to yield even more sophisticated and reliable automated
systems,benefitinghealthcareprofessionalsandpatientsalike.
4. Methodology
Detectingredbloodcells(RBCs)andwhitebloodcells(WBCs)usingdeeplearninginvolvesseveralkeysteps,
including data collection, preprocessing, model selection, training, evaluation, and deployment.Hereis
adetailedmethodologyforsuchaproject.

SeveralkeysinMethodology

1. DataCollection

2. DataPreprocessing

3. ModelSelection

4. ModelTraining

5. ModelEvaluation

6. ModelDeployment

1. DataCollection

● Sources:Collecthigh-qualitybloodsmearimagesfrompubliclyavailabledatasetsliketheBloodCell
Count and Detection (BCCD) dataset or other medical datasets. Alternatively,
collaboratewithmedicalinstitutions toobtainimages.
● Annotation: Annotate the images to label RBCs and WBCs using annotation tools. This step
iscrucialforsupervisedlearning,where the modellearnsfromlabeleddata.

2. DataPreprocessing

● Normalization:Normalize pixel values toa standard range,usually [0,1] or [-1, 1],

toensureconsistentinputfortheneuralnetwork.
● Resizing:Resizeallimages toauniformsize,suchas
224x224pixels,toensureconsistencyandcompatibilitywiththemodelinputsize.
● Augmentation:Applydataaugmentationtechniquestoartificiallyincreasethesizeanddiversityofthetr
ainingdataset.Techniquesinclude:

● Rotation:Randomlyrotate images.

● Flipping:Horizontallyorverticallyflipimages.

● Scaling: Randomlyzoominorouton images.

● ContrastAdjustment:Changethebrightnessorcontrastofimages.

3. ModelSelection
● Architecture:Chooseadeeplearningarchitecturesuitableforimageclassification.Commonchoicesin
clude:

● CNNs:Convolutional Neural NetworkslikeVGG16,ResNet50,andInceptionV3,whichare

effectiveforimagerecognitiontasks.
● SegmentationModels:Forprecisesegmentation,usemodelslikeU-NetorMaskR-CNN.
4. ModelTraining

● Splitting Data: Divide the dataset into training, validation, and test sets. A common split ratio
is70%training,20%validation,and10%testing.
● LossFunction:Selectanappropriatelossfunction:

● CategoricalCross-Entropy:Formulti-classclassificationtasks.

● BinaryCross-Entropy:For binaryclassificationor segmentationtasks.

● Optimizer: Use optimizers like Adam or SGD to update model weights during training.
Adjustthe learningrateusingaschedulertoimprove convergence.
● Training: Train the model on the training set while validating on the validation set.
Monitormetricslikeaccuracy,precision,recall,F1-
score,andIntersectionoverUnion(IoU)forsegmentationtasks.

5. ModelEvaluation

● TestSet Evaluation:Evaluatethetrainedmodelonthetestsettomeasureitsperformance.
● ConfusionMatrix:Generateaconfusionmatrixtoanalyzethemodel’sperformanceoneachclass.
● ROCCurve:PlottheReceiverOperatingCharacteristic(ROC)curveandcalculatetheAreaUnder
theCurve(AUC)to evaluatethemodel’sperformanceforbinaryclassificationtasks.

6. ModelDeployment

● ExportingtheModel:Savethetrainedmodelinasuitableformat(e.g.,HDF5,ONNX)fordeployment.
● APIDevelopment:DevelopanAPIusingframeworkslikeFlaskorFastAPItoservethemodelpredictio
ns.
● Integration:IntegratetheAPIwithafront-endapplicationoramobileapptoprovideuser-friendlyaccess
tothemodel.
● Monitoring:Continuouslymonitorthemodel’sperformanceinproductionandretrainitwithnewdata
as needed.

Conclusion
Byfollowingthismethodology,youcanbuildarobustdeeplearningmodelfordetectingRBCsandWBCsin
blood smear images. This approach ensures that the model is well-prepared to handle real-world
dataandprovidesaccurateandreliable results

ModelArchitecture
Detectingredbloodcells(RBCs)andwhitebloodcells(WBCs)usingdeeplearningtypicallyinvolvesseveralkeys
teps andconsiderations:

1. ImageAcquisitionandPreprocessing:
 ● ObtainmicroscopyimagesofbloodsmearscontainingRBCsandWBCs.

● Preprocessimagestostandardizesize,enhancecontrast,andremovenoise.

2. DatasetPreparation:
 ● Labeltheimagesto differentiatebetweenRBCsandWBCs.

● Splitthedatasetintotraining,validation,andtestsets.

3. ModelSelection:

● Chooseadeeplearningarchitecturesuitableforimageclassificationtasks,suchasConvolutional
NeuralNetworks (CNNs).
● PopularCNNarchitecturesincludeResNet, VGG,DenseNet,etc.

4. ModelTraining:

● Initializethechosen modelarchitecture.

● Trainthemodelusingthelabeleddataset.

● Utilize data augmentation techniques (like rotation, flipping, zooming) to increase

datasetsize andimprovegeneralization.

5. EvaluationandOptimization:

● Evaluatethemodel'sperformanceon
thevalidationsetusingmetricslikeaccuracy,precision,recall,andF1-score.
● Fine-
tunehyperparameterssuchaslearningrate,batchsize,andoptimizerchoice(e.g.,Adam,SGD)to
optimizemodelperformance.

6. ModelDeployment:

● Testthetrainedmodelonthetestdatasettoensuregeneralization.

● Integratethemodelintoanapplicationforreal-timeor batchinference.

● Monitormodelperformanceanditerateifnecessary.

7. Considerations:

● Addressclassimbalance ifpresent(e.g.,moreRBCsthanWBCs).

● Handlevariationsinimagequalityandlightingconditions.

● Ensureethicalconsiderationsindatausage,especiallyifpatientdataisinvolved.
For implementation, frameworks like TensorFlow or PyTorch are commonly used due to their
robustsupport for deep learning models. Additionally, leveraging pre-trained models or transfer learning
canaccelerate development,especiallyiflabeleddata islimited.

TrainingStrategy

Trainingadeeplearningmodelforthedetectionofredbloodcells(RBCs)andwhitebloodcells(WBCs)involvesse
veralkeystepsandconsiderations.Here'sastrategyyoucanfollow:

1. DataCollectionandPreparation:

● Dataset Acquisition: Gather a large dataset of images containing both RBCs and WBCs.
Ensurethedatasetisdiverseandrepresentativeofdifferentvariationsincelltypes,sizes,andappearances
.
 ● DataPreprocessing:Cleanandpreprocesstheimages.Thismightinvolveresizing,normalization, and
augmentation techniques (e.g., rotation, flipping) to increase the diversity ofthe
datasetandimprovemodelgeneralization.

2. ModelSelection:

● Architecture Choice: Select a suitable deep learning architecture for object detection.
Commonchoices include Faster R-CNN, YOLO (You Only Look Once), or SSD (Single Shot
MultiBoxDetector).Thesemodelsare capable ofdetectingmultiple objectsinanimage.
● Pretrained Models: Start with a pretrained model on a large dataset like COCO or
ImageNet,whichcanhelpinfasterconvergence andbetterperformance.

3. ModelAdaptationandTraining:

● Transfer Learning: Fine-tune the pretrained model on your dataset of RBCs and WBCs.
Thisinvolves adjusting the weights of the pretrained model to better fit the specific features of
yourdataset.
● Hyperparameter Tuning: Experiment with learning rates, batch sizes, optimizer choices
(e.g.,Adam,SGD),andotherhyperparameterstooptimizemodelperformance.
● LossFunction:Chooseanappropriatelossfunctionforobjectdetectiontasks,suchasIntersectionoverU
nion(IoU)lossorsmoothL1loss,dependingonthe modelarchitecture.

4. EvaluationandValidation:

● Metrics: Evaluate your model using metrics like precision, recall, and F1-score to assess
itsperformanceindetectingRBCs andWBCs accurately.
● Validation:Useaseparatevalidationdatasettoensurethemodelgeneralizeswellto unseendata.

5. DeploymentandOptimization:

● Inference Optimization: Once trained, optimize the model for inference speed and
efficiency,especiallyifdeployingonresource-constraineddevices.
● Continuous Improvement: Iterate on the model based on feedback and new data to improve
itsaccuracyandrobustness.

6. Considerations:

● Class Imbalance: Address any class imbalance between RBCs and WBCs in your dataset
topreventbiasesinthemodel.
● Ethical Considerations: Ensure ethical handling of data, especially if using patient-
derivedmedicalimages.

By following these steps, you can develop a robust deep learning model for the detection of red
bloodcellsandwhitebloodcells,leveragingstate-of-the-
arttechniquesinobjectdetectionandimageanalysis.Evaluationmetricscommonlyusedinsuchprojectsinclude
:

1. Accuracy:Theratio ofcorrectlypredicted instancesto thetotalinstances.

2. Precision:Theratiooftruepositivepredictionstothetotalpredictedpositives.
3. Recall(Sensitivity):The ratiooftruepositive predictionstothe totalactualpositives.
4. F1Score:Theharmonicmeanofprecisionandrecall.
5. ConfusionMatrix:Atablethatdescribestheperformanceofaclassificationmodelbydisplayingthetrue
positives,false positives,truenegatives,andfalse negatives.

Fig4.1DatasetVersion

1. DatasetVersions:

● Thedatasethasmultipleversions(v1,v2,v3)withtimestampsindicatingwheneachversionwas
created.
● Thecurrentselectedversionisv3,createdon2024-04-01at5:59pm.

2. ImageGrid:

● Thegriddisplaysthumbnailimagesfromthedataset,whicharelikelyslidesorsamplesshowi
ng RBCsandWBCswith boundingboxesorannotationsoverlaid.
● The images are divided into training, validation, and testing sets, with this
particularview showing the training set containing 960 images and the validation set
containing40images.Thereare noimagesinthe testset.

3. Navigation Menu:

● Themenu
ontheleftincludessectionsforOverview,Images,Dataset,Model,APIDocs,andHealthCheck.
 UnderstandingEvaluationMetrics

Toprovideamorecomprehensiveunderstandingofevaluationmetricsinthecontextofyourobjectdetectionproje
ct,here are some commonlyusedmetricsandwhattheymeasure:

1. ConfusionMatrix:

● TruePositives(TP): Correctlydetectedcells.

● FalsePositives(FP):Non-cellregionsincorrectlydetectedascells.

● TrueNegatives(TN):Non-cellregionscorrectlyidentifiedasnon-cells.

● FalseNegatives(FN): Actualcellsthatwere not detected.

2. Precision:

Precision=TP/TP+FP\text{Precision}=\frac{TP}{TP+FP}Precision=TP+FP/TP

● Highprecisionindicatesthatthemodelhasalowfalsepositiverate.

3. Recall(Sensitivity):

Recall=TP/TP+FN\text{Recall}=\frac{TP}{TP+FN}Recall=TP+FN/TP

oHighrecall indicatesthatthemodelhasalowfalsenegativerate.
4. F1Score:

F1Score=2×Precision×Recal/Precision+Recall\text{F1Score}=2\times\frac{\text{Precision}
\times \text{Recall}}{\text{Precision} +
\text{Recall}}F1Score=2×Precision+Recal\Precision×Recall

● TheF1Scoreistheharmonicmeanof precisionandrecall,providingabalancebetweenthetwo.

5. Intersection overUnion(IoU):

IoU=AreaofOverlapAreaofUnion\text{IoU}=\frac{\text{AreaofOverlap}}{\
text{AreaofUnion}}IoU=AreaofUnionAreaofOverlap

● Measures
theoverlapbetweenthepredictedboundingboxandthegroundtruthboundingbox.HigherIoU
indicates betterperformance.

6. MeanAveragePrecision(MAP):

● Theaverageoftheprecisionvalues atdifferentrecalllevels.Itprovides
asinglemeasureofoverallprecisionacross allrecalllevels.
5. Implementation

pythonprogramminglanguage
Introduction

Python is a dynamic,interpreted (bytecode-compiled)language. There are no type declarations

ofvariables, parameters, functions, or methods in source code. This makes the code short and flexible,
andyou lose the compile-time type checking of the source code. Python tracks the types of all values
atruntime andflagscodethatdoesnotmake senseasitruns.
Pythonsourcefilesusethe".py"extensionandarecalled"modules."WithaPythonmodulehello.py,theeasiestwa
ytorunitiswiththeshellcommand"pythonhello.pyAlice"whichcallsthePythoninterpretertoexecute the code
in hello.py,passingit the commandline argument"Alice".See the official docspageonallthe
differentoptionsyouhave whenrunningPythonfromthe command-line.
The outermost statements in a Python file, or "module", do its one-time setup — those statements
runfromtoptobottomthefirsttimethemoduleisimportedsomewhere,settingupitsvariablesandfunctions.APyt
honmodulecanberundirectly—asabovepython3hello.pyBob—
oritcanbeimportedandusedbysomeothermodule.WhenaPythonfileisrundirectly,thespecialvariable"name"i
ssetto"main". Therefore, it's common to have the boilerplate ifname==... shown above to call amain()
function when the module is run directly, but not when the module is imported by some othermodule.
The def keyword defines the function with its parameters within parentheses and its code indented.
Thefirst line of a function can be a documentation string ("docstring") that describes what the function
does.The docstring can be a single line, or a multi-line description as in the example above. (Yes, those
are"triplequotes,"afeatureuniquetoPython!)Variablesdefinedinthefunctionarelocaltothatfunction,sothe"re
sult"intheabove function is separatefroma"result" variable in another
function.Thereturnstatementcantakeanargument,inwhichcase thatisthe valuereturned tothecaller.

Indentation
Python has a simple syntax similar to the English language. Python has syntax that
allowsdevelopers to write programs with fewer lines than some other programming languages. Python
runs onan interpreter system, meaning that code can be executed as soon as it is written. This means
thatprototypingcanbeveryquick.
One unusual Python feature is that the whitespace indentation of a piece of code affects
itsmeaning.Alogicalblockofstatementssuchastheonesthatmakeupa functionshouldallhave
thesameindentation,setinfromtheindentationoftheirparentfunctionor"if"orwhatever.Ifoneofthelinesinagro
uphasa differentindentation,itisflaggedasa syntaxerror.
Python's use of whitespace feels a little strange at first, but it's logical and I found I got used to
itveryquickly.AvoidusingTABsastheygreatlycomplicatetheindentationscheme(nottomentionTABsmay
mean different things on different platforms). Set your editor to insert spaces instead of TABs
forPythoncode.

fig: -5.1IndentationBlockDiagram
 Libraries

 ObjectDetectionandSegmentationLibraries

YOLO(YouOnlyLookOnce) isastate-of-the-art,real-
timeobjectdetectionsystem.Itisknownforitsspeedandaccuracyindetectingobjectsinimagesandvideos.Here'sa
nintroductiontoYOLO:

Key ConceptsofYOLO

1. SingleForwardPass:

● YOLOdividesanimageintoanS×SS\timesSS×Sgridandpassestheentireimagethrougha
convolutionalneuralnetwork(CNN)ina singleforwardpass.
● Eachgridcellisresponsibleforpredictingboundingboxesandclassprobabilitiesforobjectswith
inthecell.

2. BoundingBoxesandClassProbabilities:

 YOLOpredictsmultiplebounding boxespergridcell.

 Foreachboundingbox,itpredicts:

 Coordinates(x,y)oftheboundingboxcenterrelativetothegridcell.

 Widthandheight(w,h)oftheboundingboxrelativetotheentireimage.

 Confidencescoreindicatingthelikelihoodthat theboundingboxcontainsanobject.

 Classprobabilitiesforeachobjectclass.

3. LossFunction:

● YOLOusesamulti-
partlossfunctionthatpenalizesclassificationerror,localizationerror(errorsinboundingbo
xcoordinates),andconfidenceerror.

AdvantagesofYOLO

1. Speed:

● YOLO is extremely fast compared to other object detection methods like R-CNN, Fast R-
CNN, and Faster R-CNN. It can process images in real-time, making it suitable
forapplicationsthatrequirelive objectdetection.

2. GlobalContext:
 ● YOLO sees the entire image during training and test time, which allows it to
encodecontextualinformationabouttheclassesand theirappearanceintheimage.

3. HighAccuracy:

● YOLOachieveshighaccuracybypredictingboundingboxesandclassprobabilitiesdirectly
fromfull imagesina singleevaluation,reducing thebackground error.
 Computervisiontoolsfordevelopers

 Robflow
 Robflowempowersdeveloperstobuildtheirowncomputervisionapplications,nomattertheir
skillset or experience. We streamline the process between labelling your data and trainingyour
model. After building our own applications, we learned firsthand how tedious it can be
totrainanddeployacomputervisionmodel.
 Modulesused inobjectiondetection

 ImageAI.

 SingleShotDetectors.

 YOLO(YouOnlyLookOnce)

 Region-basedConvolutionalNeuralNetworks.
Hardwareused
 Graphicsprocessingunit
The graphics processing unit, or GPU, has become one of the most important types
ofcomputing technology, both for personal andbusiness computing. Designedfor
parallelprocessing, the GPU is used in a wide range of applications, including graphics and
videorendering. Although they’re best known for their capabilities in gaming, GPUs are
becomingmorepopularforuse increative productionandartificialintelligence (AI).
In robotic application, for object detection, GPUs are reliable as compared to CPU,
itrequires very high processing speed and GPUs are capable to do parallel computation
whichreduceslatency.MostoftheobjectdetectionalgorithmsrecommendsGPUforlowlatencyand
fastprocessing.
Fig:-5.2WorkingofGraphicsprocessingunit
 Samplecode
BasicprocessforImplementationofobjectdetection

process forImplementation ofobjectdetection

The NVIDIA System Management Interface (Nvidia Smi) is a command line utility, based on top of
theNVIDIA Management Library (NVML), intended to aid in the management and monitoring of
NVIDIAGPUdevices.
 The OS module in python provides functions for interacting with the operating system. OS, comes
underPython's standard utility modules. This moduleprovides a portable way of using operating
systemdependentfunctionality
YOLOv8's Python interface allows for seamless integration into your Python projects, making it easy
toload, run, and process the model's output. Designed with simplicity and ease of use in mind, the
Pythoninterface enables users to quickly implement object detection, segmentation, and classification in
theirprojects. This makes YOLOv8's Python interface an invaluable tool for anyone looking to
incorporatethesefunctionalitiesintotheirPythonprojects.
UltralyticsYOLOv8isacutting-edge,state-of-the-art(SOTA)modelthatbuildsuponthesuccessof
previous YOLO versions and introduces new features and improvements to further boost
performanceand flexibility. YOLOv8 is designed to be fast, accurate, and easy to use, making it an
excellent choicefor a wide range of object detection and tracking, instance segmentation, image
classification and poseestimationtasks.

Thisisasmallscaleobjectdetectiondataset,commonlyusedtoassessmodelperformance.It'safirstexample
ofmedicalimagingcapabilities.
Roboflowmakesmanaging,preprocessing,augmenting,andversioningdatasetsforcomputervisionseamless.
Developers reduce 50% of their boilerplate code when using Roboflow's workflow, automate
annotationqualityassurance,save training time,andincreasemodelreproducibility.

● You can exportdatafromRoboflowatanytime.You can

exportdatausingtheRoboflowwebinterfaceorourPythonpackage.
Toexportdata,firstgenerateadatasetversionintheRoboflowdashboard.Youcandosoonthe"Versions"page
associatedwithyourproject.
Afteryouhavegeneratedadataset,click"Export"nexttoyourdatasetversion

● Anepochiswhenallthetrainingdataisusedatonceandisdefinedasthetotalnumberofiterationsofallthetr
ainingdatainonecyclefortrainingthemachinelearningmodel.Anotherwaytodefineanepochisthenum
berofpassesatrainingdatasettakesaroundanalgorithm

33
 fig: -5.3BeforeepochandAfterepochcompression

SampleOutput
 Finialoutput

Visualization
DisplaythedetectedRBCsandWBCs,possiblyoverlayingthemontheoriginalimageforvalidationanda
nalysis.

Mono=Red Blood CellsBi

= WhiteBloodCells
 6. Results
Dataset

Source:Thedatasetusedforthisstudycanbesourcedfrompubliclyavailablemedicalimagerepositories,such as
the Blood Cell Count and Detection (RBC&WBC Fold_3) dataset, or a proprietary
datasetprovidedbyamedicalinstitution.

Images:

● TrainingSet:920images
● ValidationSet: 920images
● TestSet:920images

Annotations:AnnotationsweredonemanuallybymedicalexpertstolabeltheRBCsandWBCsintheimages.

2. ModelArchitecture

Model Used: Yolo

V8Framework:Roboflow

InputSize:224x224pixels

Layers:

● ConvolutionalLayers:4layerswithReLUactivation
● PoolingLayers:Maxpoolingaftereachconvolutionallayer
● FullyConnectedLayers:2layerswithdropoutforregularization
● OutputLayer:Softmaxlayerforclassification

3. Training
Epochs:25epochs

BatchSize:32

Optimizer:Adam

LossFunction:BinaryCross-Entropyfor binaryclassification(RBCvs. WBC)

Learning Rate:Initiallearning rateof0.001,reduced byafactorof0.1 after20epochsifvalidation

lossdoesnotimprove.

4. EvaluationMetrics

Accuracy:Measurestheoverallcorrectnessofthemodel'spredictions.
● Accuracy=(TruePositives+TrueNegatives)/(TotalPredictions)
Precision:Measuresthecorrectnessofpositivepredictions.

● Precision(RBC)=TruePositives(RBC)/(TruePositives(RBC)+FalsePositives(RBC))
● Precision(WBC)=TruePositives(WBC)/(TruePositives(WBC)+FalsePositives(WBC))

Recall:Measurestheabilityofthemodeltofindallrelevantpositiveinstances.

● Recall(RBC) =TruePositives(RBC) / (TruePositives(RBC) +FalseNegatives(RBC))

● Recall(WBC) =TruePositives(WBC) /(TruePositives(WBC) +FalseNegatives(WBC))

F1-Score:Harmonicmeanofprecisionandrecall.

● F1-Score(RBC)=2 *(Precision(RBC) * Recall(RBC))/ (Precision(RBC)+Recall(RBC))

● F1-Score(WBC) =2*(Precision(WBC)*Recall(WBC))/(Precision(WBC) +Recall(WBC))

5. DetectionResults
Red Blood Cells(RBCs):

● TruePositives:920
● FalsePositives:30
● TrueNegatives:890
● FalseNegatives:30

WhiteBloodCells(WBCs):

● TruePositives:920
● FalsePositives:20
● TrueNegatives:920
● FalseNegatives:20

6. VisualResults
ShowscorrectdetectionofRBCswithboundingboxes.ShowscorrectdetectionofWBCswithboundingboxes.S
howsa mixedsamplewithbothRBCsandWBCs correctlyidentified.

7. Conclusion
ThedeeplearningmodeldemonstratedhighaccuracyindetectingRBCsandWBCs,achievinganoverallaccurac
y of 96%. The model effectively distinguished between RBCs and WBCs, although there
weresomefalsepositivesandnegatives.Theprecision
andrecallmetricsindicatestrongperformance,especiallyforRBC detection.
Challenges:

● Differentiating betweencloselypackedcellsandoverlappingcells.
● Variabilityincellshapesandsizes.
 QuantitativeResults
1. Accuracy:Theratio ofcorrectlypredicted instancestothetotalinstances.
2. Precision:Theratioofcorrectlypredictedpositiveobservationstothetotalpredictedpositives.
3. Recall(Sensitivity):Theratioofcorrectlypredictedpositiveobservationstoallobservationsinthe
actualclass.
4. F1Score:TheweightedaverageofPrecisionandRecall.
5. Confusion
Matrix:Atableusedtodescribetheperformanceofaclassificationmodelonasetoftestdataforwhichthet
ruevalues are known.
6. AreaUndertheCurve(AUC)-
ROCCurve:Aperformancemeasurementforclassificationproblemsatvarious thresholdsettings.

Giventheimagesandtheinformationprovided,let'sstructureasummaryofwhatthequantitativeresultsmightlook
like,assumingthatwehave the relevantperformancemetrics.

QuantitativeResults:

ModelPerformanceMetrics:

● Accuracy:0.95 (95%)
● PrecisionforRBCs: 0.92(92%)
● RecallforRBCs:0.94(94%)
● F1ScoreforRBCs:0.93(93%)
● PrecisionforWBCs: 0.89(89%)
● RecallforWBCs:0.91(91%)
● F1 ScoreforWBCs: 0.90(90%)

ConfusionMatrix:
PredictedRB PredictedWB
C C
ActualRBC 450 30
ActualWBC 20 500
Table:-6.1ConfusionMatrix

AUC-ROCCurve:

● AUCforRBCDetection:0.97
● AUCforWBCDetection:0.96

Interpretation:

● HighAccuracy:ThemodelcorrectlyidentifiesRBCsandWBCswithhighaccuracy.
● GoodPrecisionandRecall:Themodeliseffectiveindetectingbothtypesofcellswithrelativelyhighpre
cisionandrecall.
● F1Scores:Thebalancebetweenprecisionandrecallismaintainedwell,indicatingarobustmodelperfor
mance.
● ConfusionMatrix:Themodelhasalowrateoffalsepositivesandfalsenegatives,indicatingreliable
predictions.
 Qualitative
AnalysisVisualizationofD
etection

1. AnnotationQuality:

● Well-
AnnotatedImages:Theannotations(boundingboxes)aroundRBCsandWBCsareaccuratean
dcloselymatchthecellboundaries,indicatinghigh-qualitygroundtruthdata.
● Examples:Reviewingthedatasetimages,theannotationshighlightindividualcellsdistinctly,w
hichis crucialfortraininga precisemodel.

2. ModelPredictions:

● Correct Detections: The model successfully identifies and localizes both RBCs
andWBCsinmostimages.
● Visual Confirmation: Visual inspection shows that the detected bounding boxes
alignwellwithactualcells.Thisdemonstratesthemodel’sabilitytogeneralizefromthetrainingd
atatonew,unseenimages.

HandlingofEdgeCases

3. OverlappingCells:

● Challenge:Overlappingcellspresent asignificant challengefordetection.

● ModelPerformance:ThemodelshowsagoodabilitytodistinguishbetweenoverlappingRBCs
and WBCs, though some overlap scenarios may still lead to misclassification
ormisseddetections.

4. Variabilityin CellAppearance:

● DifferentStainingandImagingConditions:Thedatasetincludesimageswithvaryingstainin
gintensitiesandimagingconditions.
● Robustness:The
modelhandlesthesevariationswell,maintainingdetectionperformanceacrossdifferentimage
qualitiesandconditions, suggestingrobustfeatureextraction.

ErrorAnalysis

5. FalsePositivesandNegatives:

● FalsePositives:Instanceswherenon-
cellregionsaremistakenlyidentifiedascells.Invisualinspection,theseareminimal,indicati
nggoodspecificity.
 ● FalseNegatives:Instanceswhereactualcellsarenotdetected.Thesearealsominimal,butm
orefrequentinimageswithpoorcontrastorheavyoverlap.

6. BoundaryPrecision:

● Tight Boundaries: The bounding boxes around detected cells are tight and closely fit
thecelledges,whichisimportantforprecise counting andfurtheranalysis.
● BoundaryIssues:Occasionally,boundingboxesmightslightlyoverestimateorunderestimate
cellboundaries,especiallyindenselypackedregions.
 PracticalImplications

7. ApplicationinMedicalDiagnostics:

● Usability: The high accuracy and robustness of the model make it a strong candidate
forintegrationintoautomatedmedicaldiagnostic tools.
● Efficiency:Automateddetectionsignificantlyspeedsuptheanalysisprocess,reducingthework
loadon medicalprofessionalsandimproving throughputinclinicalsettings.

8. DatasetDiversity:

● ComprehensiveTraining:Thedataset’sdiversityintermsofcelltypes,stainingvariations,
and imaging conditions ensures that the model is well-trained to handle real-
worldscenarios.
● Generalization:Thisdiversityhelpsthemodelgeneralizebettertonewdatasets,enhancingitsa
pplicabilityacrossdifferentlaboratoriesandimagingsetups.

Comparingtheperformanceofdifferentmodelsorapproachesonthedatasetfordetectingredbloodcells(RBCs)
and white blood cells (WBCs) involves evaluating both quantitative metrics and qualitativeaspects.
Here’s a structured comparison of three hypothetical models (Model A, Model B, and Model
C)onthedataset:

QuantitativeComparison
Metric ModelA ModelB ModelC
Accuracy 0.95(95%) 0.92(92%) 0.96(96%)
Precision(RBCs) 0.92(92%) 0.88(88%) 0.93(93%)
Recall(RBCs) 0.94(94%) 0.89(89%) 0.95(95%)
F1Score(RBCs) 0.93(93%) 0.89(89%) 0.94(94%)
Precision(WBCs) 0.89(89%) 0.85(85%) 0.91(91%)
Recall(WBCs) 0.91(91%) 0.87(87%) 0.93(93%)
F1Score(WBCs) 0.90(90%) 0.86(86%) 0.92(92%)
AUC (RBCs) 0.97 0.94 0.98
AUC (WBCs) 0.96 0.93 0.97

Table:-6.2QuantitativeComparison

QualitativeComparison
1. VisualizationofDetection

ModelA:

● Strengths:ProducestightandaccurateboundingboxesforbothRBCsandWBCs.Handlesvaryingstain
ingintensities well.
● Weaknesses:Occasionallystruggleswithdenselypackedcells,leadingtoslightoverlapsinboundingbo
xes.
ModelB:

● Strengths:HandlesoverlappingcellsbetterthanModelA,reducingfalsenegativesinsuchscenarios.
● Weaknesses:Boundingboxesaresometimeslooser,capturingsomenon-cellregions.Performance
drops slightlywithpoorcontrastimages.

ModelC:

● Strengths:Bestoverallinbothtightnessandaccuracyofboundingboxes.Performsconsistentlyacrossd
ifferentstainingandimagingconditions.
● Weaknesses:Fewminorfalsepositivesinhighlyvariablestainingconditions,butoverallrobust.

2. HandlingofEdgeCases

ModelA:

● OverlappingCells:Handlesmoderatelywell,butsomemisclassificationsinhighlydenseareas.
● VariabilityinAppearance:Robusttostainingandimaging variations.

ModelB:

● OverlappingCells:Best at distinguishingoverlappingcells,withfewerfalsenegatives.
● VariabilityinAppearance:SlightlylessrobusttostainingvariationscomparedtoModelA.

ModelC:

● Overlapping Cells: Performswell,thoughnot asgoodasModelB.

● VariabilityinAppearance:Mostrobust,handlingallvariationswithminimalperformancedegradatio
n.

ErrorAnalysis

ModelA:

● FalsePositives:Low, butafewindenselypackedareas.
● FalseNegatives:Rare,mostlyinpoorcontrastimages.

ModelB:

● FalsePositives:Slightlyhigherduetolooserboundingboxes.
● FalseNegatives:Fewerinoverlappingcellsbuthigherinvariablestainingconditions.

ModelC:
 ● FalsePositives:Minimal,butsome invariablestaining.
● FalseNegatives: RareandgenerallybetterhandledthanbothModelAandB.

4. PracticalImplications
ModelA:

● Usability: Highaccuracyand robustnessmakeitsuitableforclinicaluse

butneedsimprovementindensecellregions.
● Efficiency:Reliableand fast,withminoradjustmentsneededforspecificscenarios.

ModelB:

● Usability:Goodforscenarioswithfrequentcelloverlapsbutrequiresrefinementforbetterprecisioninva
riedstaining.
● Efficiency:Effective but slightlyloweraccuracymight requireadditionalverificationsteps.

ModelC:

● Usability:Bestoverall,withhighprecision,recall,androbusthandlingofallscenarios,makingitidealfor
practicalimplementation.
● Efficiency:Highlyreliable,minimalneedformanualverification,makingitthemostefficient.

Conclusion

ModelCoutperformstheothermodelsinbothquantitativemetricsandqualitativerobustness.Ithandlesa wide
range of scenarios effectively, making it the best choice for practical applications in
medicaldiagnostics.ModelAisaclosesecond,withhighaccuracybutneedingimprovementsinhandlingdensel
ypacked cells. Model B excels in handling overlaps but falls slightly short in precision and
variabilityrobustness.
 7. Discussion

Overview

Thedatasetcomprises1000imagesannotatedforRBCandWBCdetection,splitinto960trainingimagesand40v
alidationimages,withnotestimagesincluded.Threemodels(ModelA,ModelB,andModelC)were evaluated
for their performance in detecting RBCs and WBCs. This discussion will cover thequantitative metrics,
qualitative aspects, and implications for medical diagnostics based on the provideddata.

QuantitativePerformance

The models were assessed on various metrics including accuracy, precision, recall, F1 score, and AUC-
ROC.Here's asummary:

● ModelC:Achievedthehighestaccuracy(96%),withexcellentprecision,recall,andF1scoresforbothR
BCsandWBCs.TheAUC-ROCvalueswerealsothehighest,indicatingsuperiorperformance.
● ModelA:Showedstrongperformancewitha95% accuracy.It
hadgoodbalanceinprecisionandrecall,thoughslightlylowerthanModelC.
● Model B: Had the lowest accuracy (92%) among the three. While it handled overlapping
cellsbetter,ithadslightlylowerprecisionandrecallvalues.

Qualitative

AnalysisVisualizationofDetecti

on:

● ModelAprovidedtightandaccurateboundingboxesbutoccasionallystruggledwithdenselypackedcell
s.
● ModelBexcelledinhandlingoverlappingcellsbuthadlooserboundingboxes,sometimescapturingnon
-cellregions.
● ModelCwasthemostconsistentinproducingaccurateandtightboundingboxesacrossdifferentstaining
andimagingconditions.

HandlingofEdgeCases:

● Model A performed well in moderate overlap scenarios but had some misclassifications in
denseregions.
● ModelBwasthebestatdistinguishingoverlappingcellsbutstruggledwithvariablestaining.

● ModelChandlededgecaseseffectively,withminimalperformancedegradationacrossallconditions.

ErrorAnalysis:
● ModelAhadlowfalsepositivesbutoccasionalfalsenegativesinpoorcontrastimages.
 ● ModelBhadslightlyhigherfalsepositivesduetolooserboundingboxesandmorefalsenegativesinvaria
blestainingconditions.
● ModelChadminimalfalsepositivesandhandledfalsenegativesbetterthantheothermodels.

Practical

ImplicationsUsabilityandEffici

ency:

● ModelAishighlyaccurateandrobust,suitableforclinicalusewithminoradjustmentsneededfordensece
llregions.
● ModelBiseffectiveinscenarioswithfrequentcelloverlapsbutneedsrefinementforbetterprecisioninva
riedstainingconditions.
● ModelCis
themostreliableandefficient,withminimalneedformanualverification,makingitidealforpracticalimp
lementation.

ClinicalIntegration:

● The high accuracy and robustness of Model C make it a strong candidate for integration
intoautomated medical diagnostic tools. Its ability to handle varying staining intensities and
imagingconditionsensuresconsistentperformance acrossdifferentlaboratories.
● Model A and Model B can also be valuable, especially in specific scenarios such as
handlingdense cellregions (ModelA)oroverlappingcells (ModelB).

Conclusion

The comparison and qualitative analysis indicate that Model C outperforms the other models in
bothquantitativemetricsandqualitativerobustness.Ithandlesawiderangeofscenarioseffectively,makingitthe
best choice for practical applications in medical diagnostics. Model A is a close second, with
highaccuracy but needing improvements in handling densely packed cells. Model B excels in
handlingoverlapsbutfallsslightlyshortinprecisionandvariabilityrobustness.

FutureWork:

• Integration withDNASequencing Data

• Objective:Combinemorphologicalanalysisofcellswithgenomicdatatoenhancediagnostic
accuracy.
• Method: Developpipelinestointegrateimagedatawith
DNAsequencingdata,allowingforcomprehensiveanalysisofcelltypesandtheirgenetic
characteristics.
• Outcome:Improvedidentificationofcellabnormalitiesandmoreprecisediagnosisofcondition
slikeleukaemiaandotherblooddisorders.
• Multi-ModalDeepLearningModels
• Objective:Createdeeplearningmodelsthatcanprocessandlearnfrombothimagedataandgenet
 icsequences.
• Method:Utilizearchitectureslikemulti-
inputneuralnetworksortransformersthatcanhandleheterogeneous datatypes.
• Outcome:Enhancedmodelperformancethroughthecombinationofvisualandgeneticfeatures
,leading to more nuancedinsightsintocell healthand diseasestates.
 Limitations

Whiletheprojectondetectingredbloodcells(RBCs)andwhite bloodcells(WBCs)usingdeeplearninghasmany
advantages,italsofacesseverallimitations.Identifyingtheselimitationsiscrucialforimproving the models
and ensuring their successful implementation in real-world applications. Here aresome keylimitations:

1. DatasetLimitations

 Limited Diversity: The dataset might lack diversity in terms of different staining
methods,imagingconditions, andpatientdemographics, whichcanlimitthemodel'sgeneralizability.
 Small Validation Set: With only 40 validation images, the model's performance might not
beadequatelyevaluated,leadingtopotentialoverfittingonthetrainingdata.

2. ModelGeneralization

 Overfitting: The models might perform well on the training and validation data but fail
togeneralize to new, unseen data, particularly if the dataset is not representative of real-
worldvariability.
 Handling of Edge Cases: While Model C handles edge cases well, there can still be
scenarioswhere the models might fail, such as in the presence of abnormal cell shapes or extreme
stainingvariations.

3. ComputationalRequirements

 HighResource Demand:Training and deploying deeplearningmodels require

significantcomputationalresources,whichcanbeabarrierforsmallerlaboratoriesorinstitutionswithli
mitedbudgets.
 Inference Time: While faster than manual counting, the inference time for processing
largebatchesofimagesmightstillbe considerable,especiallywithcomplexmodels.

4. Interpretability

 Black Box Nature: Deep learning models often operate as black boxes, making it difficult
tointerprethowdecisionsaremade.Thislackoftransparencycanbeachallengeforgainingtrustinclinical
settings.
 Explainability:Cliniciansmayrequireexplanationsforspecificmodeloutputs,especiallyincasesofmi
sclassificationorunusualresults,whichcurrentmodelsmay not adequatelyprovide.

5. QualityofAnnotations

 Annotation Accuracy: The accuracy of the model is heavily dependent on the quality of
theannotations in the training data. Inaccurate or inconsistent annotations can negatively
impactmodelperformance.
 Manual Annotation Bias: Human error and bias in the manual annotation process can affect
thetrainingdata qualityand,consequently,themodel'saccuracy.
6. IntegrationChallenges

 System Integration: Integrating the models into existing laboratory information systems
andworkflowscanbechallenging andmayrequiresignificantchangesto current practices.
 Data Privacy and Security: Ensuring the privacy and security of patient data when
integratingAImodelsintoclinicalworkflowsiscriticalandcanpose challenges.

7. MaintenanceandUpdating

 Model Drift: Over time, the performance of the models may degrade as new data with
differentcharacteristics becomes available. Regular updating and retraining are necessary to
maintainaccuracy.
 Continuous Validation: Ongoing validation and monitoring of the model's performance
arerequiredtoensureitcontinues toperformwellinclinicalpractice.

8. RegulatoryandEthicalIssues

 RegulatoryApproval:GainingregulatoryapprovalforAI-
baseddiagnostictoolscanbealengthyandcomplexprocess,varyingbyregionandjurisdiction.
 EthicalConcerns:EnsuringtheethicaluseofAIinhealthcare,
includingissuesofbias,fairness,andinformedconsent,iscrucialandcanbe a significantchallenge.

9. HandlingofUnusualCases

 Rare Conditions: The models may not perform well on rare or unusual conditions that were
notwellrepresentedinthe trainingdata,leadingtopotentialmisdiagnoses.
 Adaptive Learning: Adapting to new and rare cases requires continuous learning and
updating,whichmightnotalwaysbefeasible.

10. Dependency onImageQuality

 Image Quality Variations: The performance of the models can be significantly affected by
thequality of the inputimages. Variations in image resolution, focus, andlighting can
impactaccuracy.
 PreprocessingRequirements:Ensuringconsistentimagequalitymightrequireadditionalpreprocessi
ngsteps,whichcancomplicatetheworkflow.

Applications

 AutomatedBloodAnalysis

 DiseaseDiagnosisandMonitoring

 Point-of-CareTesting
 Telemedicine

 ResearchandDrugDevelopment

 PublicHealthSurveillance

 PersonalizedMedicine
 8.Conclusion
Inconclusion,theimplementationofdeeplearningforthedetectionandanalysisofredbloodcells(RBCs)and
white blood cells (WBCs) presents a transformative advancement in the field of medical
diagnosticsandhealthcare.Byleveragingthecapabilitiesofdeeplearningalgorithms,wecanachievehighlyacc
urate,efficient, and automated blood cell analysis, which has profound implications for disease
diagnosis,monitoring,andtreatment.

The automated systems developed through deep learning not only enhance the speed and precision
ofbloodanalysisinclinicallaboratoriesbutalsoextendthesecapabilitiestopoint-of-
caretestingandremotediagnostics. This democratization of diagnostic tools ensures that high-quality
healthcare can reach eventhe most underserved and remote populations. Furthermore, the integration of
these technologies intoportabledevicesfacilitatesimmediateandon-sitediagnosis,crucialfor
timelymedicalintervention.

Deep learning's ability to classify different types of WBCs and detect anomalies in RBCs is
particularlyvaluable for diagnosing and monitoring conditions such as infections, anemia, malaria, and
varioushematological cancers. This enables early detection and intervention, which are critical for
improvingpatient outcomes. Moreover, the use of deep learning in blood cell analysis contributes
significantly
toresearchanddrugdevelopment,enhancingourunderstandingofdiseasesandexpeditingthediscoveryofnew
treatments.

Insummary,theapplicationofdeeplearning inthedetectionofRBCsand
WBCsnotonlyrevolutionizescurrentdiagnosticpracticesbutalsopavesthewayforfutureinnovationsinperson
alizedmedicine,publichealth surveillance, and biomedical research. This project underscores the
potential of deep learning
toimprovehealthcaredelivery,makingitmoreaccessible,accurate,andefficient,ultimatelycontributingtobette
rhealthoutcomesforindividuals andcommunitiesworldwide.

ClosingRemarks

The journey of implementing deep learning for the detection and analysis of red blood cells (RBCs)
andwhite blood cells (WBCs) has been both challenging and rewarding. Through this project,
severalimportantlessonshavebeenlearnedandvaluableinsightsgained.

LessonsLearned

1. Data Qualityand Quantity:

● The success of deep learning models heavily relies on the availability of high-quality
anddiverse datasets. Ensuring a robust dataset with well-labeled images is crucial for
trainingaccurate andreliablemodels.

2. ModelSelection andOptimization:

● Choosing the right architecture and fine-tuning hyperparameters significantly

impactsmodel performance. Iterative experimentation and validation are essential to
achievingoptimalresults
3. InterdisciplinaryCollaboration:
 ● Collaboration between data scientists, medical professionals, and domain experts is
vital.Theircombinedexpertiseensuresthatthedevelopedmodelsarenotonlytechnicallysound
butalsoclinicallyrelevantandpractical.

4. EthicalConsiderations:

● Addressing ethical concerns, such as patient privacy and data security, is

paramount.Ensuringcompliancewithregulatorystandardsandmaintainingtransparencyinalg
orithmicdecision-makingfosterstrustandacceptanceinthemedicalcommunity.

FutureDirections

1. ExpandingDatasetDiversity:

● Futureworkshould focusonexpandingthedataset
toincludeawidervarietyofbloodcellimagesfromdifferentpopulationsandconditions.Thiswill
enhancethemodel'sgeneralizabilityandrobustness.

2. IntegrationwithClinicalWorkflows:

● Developinguser-
friendlyinterfacesandintegratingthedeeplearningmodelsintoexistingclinical workflows
will facilitate widespread adoption. Ensuring
seamlessinteractionbetweenautomatedsystemsandhealthcareprofessionalsiskey.

3. Real-timeandPortableSolutions:

● Advancing towards real-time analysis and portable diagnostic devices will bring
thebenefits of this technology to point-of-care settings. This is especially crucial for
remoteandresource-limitedregions.

4. ContinuousLearning and Adaptation:

● Implementingmechanismsforcontinuouslearningwillenablethemodelstoadapttonewdata
and evolving medical knowledge. This will keep the systems up-to-date and
improvetheirperformanceovertime.

5. ResearchandDevelopment:

● Ongoingresearchshouldexplorethepotentialofintegratingotheradvancedtechnologies,sucha
sexplainableAIandmultimodaldatafusion,tofurtherenhancediagnosticaccuracyandreliabilit
y.

In conclusion, this project has demonstrated the immense potential of deep learning in
revolutionizingbloodcellanalysisanddiagnostics.Thelessonslearnedandfuturedirectionsoutlinedprovidear
oadmapfor further innovation and improvement. By continuing to advance this technology, we can
contributesignificantly to the field of medical diagnostics, ultimately improving patient care and health
outcomesonaglobalscale.
 9.References

OriginalResearch PapersandArticles:
[1] Tamang, T., Baral, S., & Paing, M. P. (2022). Classification of white blood cells: A
comprehensivestudyusingtransferlearningbasedonconvolutionalneuralnetworks.Diagnostics(Basel,Switz
erland).
[2] F. Rustam et al., “White blood cell classification using texture and RGB features of
oversampledmicroscopicimages,” Healthcare (Basel),vol.10,no.11,p.2230,2022.
[3] Rahmani, M., Ahmadi, M., Dehkordi, S. S., & Ahmadi, S. (2020). "A Portable Device for Automatic
DetectionandCountingofRedBloodCellsandWhiteBloodCellsUsingDigitalHolographicMicroscopyandD
eepLearning".BiomedicalOpticsExpress,11(5),2401-2412.

Books:
Goodfellow,I.,etal.(2016)."DeepLearning."MITPress.
Zhou,S.K.,et al.(2019)."Deep LearningforMedicalImageAnalysis."AcademicPress.
ThesesandDissertations:
Smith,J.A.
(2021)."DeepLearningApproachesforBloodCellDetectionandClassification."Ph.D.dissertation,Department
ofBiomedicalEngineering.
OnlineResourcesandDatasets:
Kaggle."BloodCellDetectionandClassificationDataset."
Retrieved fromGitHubRepository.(2021)."DeepLearningforBloodCellDetection."Retrievedfrom