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.
