WBC-Net: A Convolutional Neural Network for Automated White Blood Cell Classification

WBC-Net is a deep convolutional neural network for automated classification of white blood cells.
Neutrophils, Lymphocytes, Monocytes, Eosinophils, and Basophils.

---

Abstract

White blood cell (WBC) analysis is crucial for early detection of hematological disorders.
This repository provides the implementation of WBC-Net, a convolutional neural network designed for robust and scalable WBC classification from microscopic images.

The pipeline includes advanced image preprocessing (resizing, normalization, and augmentation), a four-stage CNN with batch normalization and dropout regularization, and performance optimization techniques such as early stopping, learning-rate scheduling, and model checkpointing.
Extensive experiments on a public WBC dataset show that WBC-Net achieves strong performance across multiple metrics, making it suitable for real-time clinical decision-support applications.

---

Table of Contents

1. Project Overview
2. Dataset
3. Methodology
4. Experiments & Results
5. Repository Structure
6. Installation
7. Usage
8. Citation
9. Contact

---

1. Project Overview

The goal of WBC-Net is to automate multi-class WBC classification using a deep learning pipeline optimized for accuracy, robustness, and reproducibility.

Key Contributions

• A 4-stage CNN architecture tailored for WBC morphology
• Comprehensive preprocessing pipeline including resizing, augmentation, and normalization
• Training optimization: batch normalization, dropout, learning-rate scheduling
• Evaluation suite: confusion matrix, ROC curves, per-class metrics
• GPU-optimized implementation

---

2. Dataset

• Source:
  White Blood Cells Dataset

• Classes (5):

  • Neutrophils
  • Eosinophils
  • Basophils
  • Monocytes
  • Lymphocytes

• Total Images: 14,514

  • Training: 8,140 (80%)
  • Validation: 2,035 (20%)
  • Test: 4,339

---

3. Methodology

3.1 Preprocessing Pipeline

• Resize images to 128×128
• Pixel normalization
• Augmentation: rotations, flips
• Addressing class imbalance

3.2 CNN Architecture

Input (128x128x3)
├── Conv Block 1 (64 filters) → BatchNorm → MaxPool → Dropout(0.25)
├── Conv Block 2 (128 filters) → BatchNorm → MaxPool → Dropout(0.25)
├── Conv Block 3 (256 filters) → BatchNorm → MaxPool → Dropout(0.25)
├── Conv Block 4 (512 filters) → BatchNorm → MaxPool → Dropout(0.25)
├── Dense(1024) → Dropout(0.5)
├── Dense(512) → Dropout(0.5)
└── Output(5) → Softmax

---

4. Experiments & Results

Confusion Matrix

Class Distribution

Sample Prediction

Visualizing Model Predictions

---

5. Repository Structure

WBC-Net/
 ├── models/
 ├── figures/
 ├── data/
 ├── notebooks/
 ├── scripts/
 └── README.md

---

Installation

git clone https://github.com/YOUR_USERNAME/WBC-Net.git
cd WBC-Net
pip install -r requirements.txt

---

Usage

Training

python train.py --dataset data/

Evaluation

python evaluate.py --weights best_model.h5

---

Citation

If you use WBC-Net in your research, please cite:

Ahmed T. Mersal et al.   
WBC-Net: A Convolutional Neural Network for Automated White Blood Cell Classification.  
TechRxiv, 2025.  
DOI: 10.36227/techrxiv.176282158.83951263/v1

---

Contact

Ahmed Talaat
📧 ahmed.mmt3411@gmail.com