Making sense of MRI
Exploring multi-region brain tumor segmentation with multimodal MRI and an attention-based U-Net.
PyTorch / U-Net / Medical imaging
The idea
Different MRI modalities reveal different information. This project explores how to combine them for multi-region brain tumor segmentation.
How I built it
I built a 3D preprocessing workflow with modality-wise Z-score normalization, background cropping to 128³, label remapping, and disk-cached transforms. A 2D Multi-Scale Attention U-Net uses an ImageNet-pretrained EfficientNetB3 backbone and three stacked MRI modalities.
Where it landed
An implementation covering preprocessing and the segmentation architecture. This is a research project; it is not a clinical diagnostic tool.