rehabaam/ds-cxr-covid19

Chest X-Ray Classification Model (🦠)

📋 Overview

This project focuses on building and evaluating a Convolutional Neural Network (CNN) model for classifying chest X-ray images into four categories:

• Normal
• Pneumonia
• Lung Opacity
• COVID-19

The model was trained using masked chest X-ray images (lungs only) to enhance focus on medically relevant areas.

⸻

🧠 Model Architecture

The CNN model includes:

• Input size: (256, 256, 1) RGB masked lung images
• Convolutional blocks: Conv2D(32) → Conv2D(64) → Conv2D(128) → Conv2D(256) → Conv2D(512)
• ASPP Block: Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale features.
• Attention Block: Squeeze-and-Excitation (SE Block) applied after key stages.
• Pooling Layers: Global Average Pooling 2D
• Custom Loss function: Focuses more on hard examples and less on easy one.
• Classifier Head: Dense → Softmax for multiclass classification (4 classes)

Additional techniques used:

• Data Augmentation: Random flipping, rotation (range from 0 to 10 degrees)
• Dropout: Regularization to prevent overfitting (20%)
• EarlyStopping & ReduceLROnPlateau: For efficient training

📊 Metrics

Final evaluation results:

Metric	Score
Accuracy	~93%
Precision	~92%
Recall	~92%
F1-Score	~92%

Note:

• The dataset was balanced manually into training and validation datasets (80%/20%)
• Grad-CAM visualization was used to verify model attention inside the lungs.
• The model is still being improved for higher F1 scores.

🗃 Dataset

• Source: https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database
• Masked lungs: Masked lungs were generated using GAN model (maja011235/lung-segmentation-gan)

🚀 Future Work

• Fine-tuning with different loss functions
• Model ensembling
• Clinical-grade evaluation with external datasets