brain-mri-segmentation

Production-grade binary semantic segmentation of brain tumors from MRI scans - pixel-level mask prediction for low-grade glioma (LGG) regions.

Overview

Task	Binary semantic segmentation (tumor vs. background)
Dataset	Mateusz Buda LGG MRI (TCGA) - 110 patients, 3 929 paired slices
Main model	SegFormer-B2 (nvidia/segformer-b2-finetuned-ade-512-512, ~25 M params)
Baseline	Small U-Net (4 levels, 32→256 ch, ~1.9 M params, hand-rolled)
Stack	PyTorch Lightning · Hydra · MLflow · DVC · FastAPI · Docker · GitHub Actions · MkDocs
License	MIT

Visualizations

Qualitative results on held-out test slices - input FLAIR, ground-truth mask, and SegFormer-B2 prediction side by side:

Test-set metric comparison between the SegFormer-B2 main model and the U-Net baseline (Dice, IoU, pixel accuracy):

Sections

• Architecture - data flow, model choices, metrics rationale
• Training - running experiments, logging, overrides
• Serving - FastAPI endpoints, Docker deployment
• Benchmarks - vs literature, trade-offs
• Reproducibility - pinned environment, one-command re-run
• Limitations - failure modes, dataset bias
• Model card - HF Hub card template

Links

• Code: GitHub
• Model: Hugging Face

Disclaimer

Research/educational artifact only - not intended for clinical use.