Leukemia-Diagnosis-EfficientNetV2-vs-CNNs

Overview

This project implements and compares two deep learning models for automated detection and classification of leukemia cells from microscopic images:

• Enhanced EfficientNetV2-S Model: A transfer learning approach using a pre-trained model with custom classification layers.
• Traditional CNN Model: A lightweight convolutional neural network built from scratch.

Dataset

The project uses the ALL (Acute Lymphoblastic Leukemia) dataset containing 3,256 cell images across four classes:

• Benign: 504 images (normal lymphocytes)
• Early: 985 images (early-stage leukemia cells)
• Pre: 963 images (pre-cancerous leukemia cells)
• Pro: 804 images (progressive/proliferative leukemia cells)

Leukemia Types: (A) Benign, (B) Early-stage, (C) Pre-cancerous, (D) Progressive (Proliferative)

Installation

1. Clone the repository:

   git clone https://github.com/divyanshupatel17/Leukemia-Diagnosis-EfficientNetV2-vs-CNNs.git

2. Navigate to the project directory:

   cd Leukemia-Diagnosis-EfficientNetV2-vs-CNNs

3. Install the required dependencies:

   pip install -r requirements.txt

Project Structure

├── data/                          # Dataset directory
│   ├── ALL_dataset/               # Original ALL dataset
│   │   ├── Benign/                # 504 normal lymphocyte images
│   │   ├── Early/                 # 985 early-stage leukemia cell images
│   │   ├── Pre/                   # 963 pre-cancerous leukemia cell images
│   │   └── Pro/                   # 804 progressive leukemia cell images
│   ├── cnn_model/                 # Processed data for CNN model
│   └── efficientnet_model/        # Processed data for EfficientNetV2
├── docs/                          # Documentation
│   ├── assets/                    # Project assets
│   │   ├── figures/               # Model diagrams and plots
│   │   ├── tables/                # Performance metrics tables
│   └── report/                    # Project reports and papers
│       ├── A1F2_10_REFERENCES_PAPER/   # Reference papers used in the project
│       ├── A1F2_10_SOURCE_FILE_LINKS/  # Links to source files and resources
│       ├── A1F2_10_FINAL_PAPER/        # Final project report
│       └── A1F2_10_BASEPAPER/          # Base research papers for the project
├── models/                        # Model implementations
│   ├── cnn_model/                 # Traditional CNN implementation
│   │   ├── config.py             # Model configuration
│   │   ├── models/               # Model architecture
│   │   ├── scripts/              # Training and evaluation scripts
│   │   └── results/              # Training results and metrics
│   ├── efficientnet_model/       # EfficientNetV2 implementation
│   │   ├── config.py             # Model configuration
│   │   ├── models/               # Model architecture
│   │   ├── scripts/              # Training and evaluation scripts
│   │   └── results/              # Training results and metrics
│   └── saved_models/             # Trained model weights
│       ├── efficientnetv2s_leukemia_model.h5  # Trained EfficientNetV2 model
│       └── cnn_leukemia_model.h5              # Trained CNN model
├── utils/                         # Utility functions
│   ├── compare_models.py         # Model comparison tools
│   ├── find_sample_images.py     # Image sampling utilities
│   └── test_model_loading.py     # Model loading tests
├── visualization/                 # Visualization tools
│   ├── gradcam/                  # Grad-CAM implementation
│   ├── gradcam_figures/          # Grad-CAM visualizations
│   └── model_diagrams/           # Model architecture diagrams
├── requirements.txt              # Project dependencies
└── run_guide.py                  # Interactive execution guide

Usage

1. Run the interactive guide to execute different components:

   python run_guide.py

2. The guide provides options for:
   • Training models
   • Evaluating model performance
   • Generating visualizations
   • Comparing model results

Project Results

Model Performance Comparison

Model	Accuracy	Precision	Recall	F1-Score	Training Time
EfficientNetV2-S	98.93%	98.95%	98.93%	98.93%	1.5 hours
Traditional CNN	89.42%	89.55%	89.42%	89.40%	45 minutes

Key Findings

1. Superior Transfer Learning Performance: The EfficientNetV2-S model with transfer learning significantly outperformed the traditional CNN in all metrics, demonstrating the effectiveness of leveraging pre-trained weights for medical image classification tasks.

2. Class-wise Performance: Both models performed best on the "Benign" class with nearly perfect accuracy, while "Early" and "Pre" classes were more challenging to distinguish due to their visual similarities.

3. Resource Efficiency: Despite its superior performance, the EfficientNetV2-S model required more computational resources and training time compared to the traditional CNN.

4. Inference Speed: The traditional CNN model, while less accurate, offers faster inference times, making it potentially suitable for resource-constrained environments.

5. Attention Visualization: Grad-CAM visualizations revealed that the EfficientNetV2-S model focuses on clinically relevant features in the cell images, aligning with pathological criteria used by hematologists.

Model Details

Enhanced EfficientNetV2-S

• Transfer learning with pre-trained EfficientNetV2-S
• Custom classification layers
• Achieved 98.93% accuracy
• Superior feature extraction capabilities

Traditional CNN

• Custom-built lightweight architecture
• Achieved 89.42% accuracy
• Baseline model for comparison

Visualization Tools

Grad-CAM Visualization

• Implemented in visualization/gradcam/
• Reveals model attention patterns
• Helps validate model decisions
• Shows correlation with pathological features

Model Architecture Diagrams

• Available in visualization/model_diagrams/
• Detailed network architecture visualization
• Layer-by-layer model structure

Performance Metrics

Detailed performance metrics are available in docs/assets/tables/:

• Classification accuracy by class
• Confusion matrices
• Precision, recall, and F1-scores
• Comparative analysis with existing approaches

Documentation

The docs directory contains comprehensive documentation:

• docs/assets/figures: Visual representations of model architectures and results
• docs/assets/tables: Detailed performance metrics and comparative analysis
• docs/report: Research papers, references, and project documentation including:
  • Final project report
  • Base papers for reference
  • Related works in the domain
  • Source file links for additional resources

Contact

For questions or feedback, please contact Divyanshu Patel.

• LinkedIn: linkedin.com/in/patel-divyanshu