Neurofusion: 3D Brain Tumor Segmentation
85%+ Dice · 1st Place COMPPEC 2025
Final year project: segmenting brain tumor regions from multi-modal MRI (T1-CE, T2, FLAIR) on the BraTS20 dataset. Benchmarked 3D U-Net against SegFormer3D, selected SegFormer3D, and used GradCAM for explainability and FastAPI for serving.
Problem
Segment brain tumor regions from multi-modal MRI (T1-CE, T2, FLAIR) accurately, and in a way clinicians can inspect.
Approach
- BraTS20 dataset with multi-modal MRI inputs
- Benchmarked 3D U-Net against SegFormer3D and selected SegFormer3D
- GradCAM for explainability and FastAPI for serving
Demo
Results
- 85%+ Dice score
- 1st Place at COMPPEC 2025
- Departmental Silver Medal
Stack
- PyTorch
- SegFormer3D
- 3D U-Net
- GradCAM
- FastAPI
By Muzammil Nawaz Khan, AI Engineer.