chest-xray-classifier

Chest X-Ray Classification (Normal / Pneumonia / Tuberculosis)

Project Overview

This project adapts and improves upon a Kaggle baseline to create a robust 3-class Chest X-Ray classification model.

Datasets Used

The model was trained on a merged and deduplicated dataset of over 14,000 images from:

• Kaggle jtiptj
• Rahman et al.
• Shenzhen
• Mendeley Pakistan

Model Architecture

• Architecture: ResNet101 (PyTorch)
• Fine-Tuning: Pre-trained on ImageNet, with partial fine-tuning applied to layer4 and the fully connected layer.
• Data Augmentation: Rotation, Color Jitter (Brightness), Affine Translations, Horizontal Flips.

Performance

The final model achieved the following metrics on the held-out validation and test sets:

• Validation Accuracy: 97.90%
• Test Accuracy: 98.36%

Usage

Simply upload a grayscale Chest X-Ray image to the Gradio interface to receive real-time probability predictions for Normal, Pneumonia, and Tuberculosis.