Brain Tumor Segmentation and Explainability Pipeline

A personal project, built before St. Jude, to work through what a clinical imaging pipeline looks like end to end: raw scans in, an interpretable result out.

The data is BraTS 2021: 1,251 patients with four MRI modalities each (T1, T1ce, T2, FLAIR) and expert tumor masks. The curation stage z-score normalizes each modality over brain tissue only, keeps the middle 60% of axial slices, and writes a fixed 70/15/15 train/validation/test split to a manifest so every run uses the same patients.

On top of that sits a 2D U-Net in PyTorch. The baseline, trained on a 100-patient subset, reaches a test Dice of 0.89. Most failures are small lesions the model misses entirely or leaks past at the border. Around the model are Monte Carlo dropout for slice-level uncertainty, Grad-CAM overlays, per-patient quantification (volume in mm³, bounding box, shape features), and a small Streamlit app for looking at a single case.