TBX11K Faster R-CNN (ResNet50-FPN-V2)
This model is a Faster R-CNN detector fine-tuned on the TBX11K chest X-ray dataset for Tuberculosis localization and categorization.
Training Configuration
- Image Size: 512x512
- Epochs: 30
- Batch Size: 4
- Optimization Strategy: WeightedRandomSampler for class balance control
Validation Performance (Best Epoch)
| Metric | Score |
|---|
| Best Validation mAP@0.5 | 0.3728 |
Blind Test Inference Distribution
The model was run on 3302 unannotated test samples, outputting the following diagnosis breakdown:
- Predicted Active TB: 420
- Predicted Latent TB: 60
- Predicted Healthy/Negative: 2822
Internal Model Class Map (Background Offset)
0: Background (Internal Only)1: Active TB2: Latent TB
