COVID-19 CT Scan Classification DenseNet PyTorch

📋 Table of Contents

• Overview
• Dataset Information
• Model Architecture
• Performance Metrics
• Data Visualization
• Installation & Usage
• Results & Analysis
• Future Improvements
• References

🎯 Overview

This project implements deep learning models for automated COVID-19 detection from CT scan images using PyTorch. The notebook demonstrates two approaches:

1. Custom CNN Architecture - Achieving 92% accuracy
2. DenseNet Transfer Learning - Achieving 96% accuracy

Key Features

• ✅ Binary classification (COVID-19 vs Non-COVID-19)
• ✅ Transfer learning implementation
• ✅ Data augmentation techniques
• ✅ Model comparison and evaluation
• ✅ Visualization of predictions and feature maps

📊 Dataset Information

SARS-CoV-2 CT Scan Dataset

The dataset contains CT scan images collected from real patients:

Dataset Structure:
├── COVID-19/          # CT scans from COVID-19 positive patients
│   └── ~1,252 images
└── Non-COVID/         # CT scans from healthy or other pneumonia patients
    └── ~1,229 images

Total Images: ~2,481 CT scans
Image Format: PNG
Image Size: Variable (resized to 224x224 for training)

Dataset Statistics

Metric	Value
Total Images	2,481
COVID-19 Cases	1,252 (51%)
Non-COVID Cases	1,229 (49%)
Training Set	2108 (85%)
Validation Set	186 (7.5%)
Test Set	187 (7.5%)

Class Distribution

Class / Set	Training	Validation	Test
COVID-19 Cases	1055	100	97
Non-COVID Cases	1053	86	90

📈 Performance Metrics

Model Comparison

Model	Accuracy	Precision	Recall	F1-Score
Custom CNN	92.0%	92.0%	92.0%	92.0%
DenseNet-69	96.0%	96.0%	96.0%	96.0%

DenseNet-69 Detailed Metrics

Classification Report:

Confusion Matrix

📊 Data Visualization

Training History

Accuracy Over Epochs:

Loss Over Epochs:

🔧 Installation & Usage

Installation

# Clone the repository
git clone https://github.com/yourusername/covid19-ct-classification.git
cd covid19-ct-classification

# Create Virtual Environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

📉 Results & Analysis

Key Findings

1. DenseNet Outperforms CNN

   • 4% accuracy improvement over custom CNN
   • Better generalization on test set
   • Fewer false negatives (critical for medical diagnosis)

2. Transfer Learning Benefits

   • Faster convergence (15 epochs vs 35 epochs)
   • Better feature extraction from pre-trained weights
   • More robust to small dataset size

3. Error Analysis

   • Most false positives: Early-stage COVID cases
   • Most false negatives: Atypical CT presentations
   • Confusion mainly in borderline cases

Training Configuration

Hyperparameters:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Custom CNN:
  • Optimizer: Adam
  • Learning Rate: 5e-5
  • Batch Size: 16
  • Epochs: 50

DenseNet-69:
  • Optimizer: Adam
  • Learning Rate: 3e-5
  • Batch Size: 16
  • Epochs: 50 finished 27
  • Weight Decay: 0.01
  • LR Scheduler: ReduceLROnPlateau

Data Augmentation

Training Augmentations:
• Random Horizontal Flip (p=0.5)
• Random Vertical Flip (p=0.5)
• Random Rotation (±10°)
• Normalization (ImageNet stats)

🚀 Future Improvements

Short-term Goals

• Implement ensemble methods (CNN + DenseNet)
• Add Grad-CAM visualizations for interpretability
• Experiment with other architectures (EfficientNet)
• Increase dataset size through additional sources

Long-term Goals

• Multi-class classification (COVID vs Pneumonia vs Normal)
• 3D CNN for volumetric CT analysis
• Real-time deployment as web application

📚 References

Datasets

• SARS-CoV-2 CT Scan Dataset: Kaggle

Related Medical Projects

• GANs for Kidney and Brain Tumors
• MRI Brain Tumor Classification
• Medical Text Classification using BiLSTM, BiGRU & Conv1D
• ECG Signal Classification using RNN, GRU & LSTM

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📞 Contact

For questions or collaboration opportunities:

• Kaggle: ahmedashrafhelmi
• Project Link: COVID-19 CT Classification Notebook