ViT Fine-Tuned for Chest X-Ray Pneumonia Detection
Fine-tuned from google/vit-base-patch16-224-in21k on the
Chest X-Ray Images (Pneumonia) dataset.
Research Question
Given the exact same backbone, how close can a model fine-tuned from scratch get to
nickmuchi/vit-finetuned-chest-xray-pneumonia?
Test Set Results
| Metric | nickmuchi (reference) | This model |
|---|---|---|
| Accuracy | 0.9535 | 0.9199 |
| AUC-ROC | 0.9916 | 0.9842 |
| F1 | 0.9630 | 0.9392 |
Training Details
- Backbone:
google/vit-base-patch16-224-in21k - Phase 1: Feature extraction (frozen backbone), 4 epochs, LR=3e-4
- Phase 2: Full fine-tuning, 6 epochs, LR=2e-5
- Class weights: Applied to handle ~3:1 Pneumonia:Normal imbalance
- Augmentations: Random horizontal flip, ±10° rotation, brightness/contrast jitter
Disclaimer
For educational and portfolio purposes only. Not validated for clinical use.
- Downloads last month
- 9