- Designed an end-to-end framework for segmenting organs in the head and neck CT images, achieving near-SOTA performance on the MICCAI Auto-segmentation Challenge (72.16 vs. 75.28 dice score).
- Enhanced the baseline U-Net architecture by adding residual connections and a mixup training regime to prevent overfitting on the relatively small dataset (only ~30 CT scans).
- Used Grad-CAM to interpret and understand the model predictions better.
