NeuroTrace
Explainable deep learning for brain-tumor segmentation
NeuroTrace explores brain-tumor segmentation on the BraTS 2023 glioma dataset. The principal architecture is a 3D Attention U-Net with GNN-driven attention at the bottleneck, supported by qualitative explainability and uncertainty work.
Context
The problem space.
Brain-tumor segmentation is spatially complex and clinically sensitive. The work focuses on a volumetric segmentation pipeline while also making model attention and uncertainty more inspectable.
Confirmed work
What the evidence supports.
- A 3D Attention U-Net forms the principal segmentation architecture.
- The bottleneck uses GNN-based, GNN-driven attention rather than Transformer attention.
- Inputs use patch cropping and z-score normalization on the BraTS 2023 glioma dataset.
- Training used combined Cross-Entropy and Dice loss with Adam at a learning rate around 1e-4.
- Grad-CAM and Monte Carlo Dropout support qualitative explainability and uncertainty exploration.
System flow
- 01
BraTS 2023 volumes
- 02
Patch cropping
- 03
Z-score normalization
- 04
3D Attention U-Net
- 05
GNN-driven bottleneck attention
- 06
Segmentation output
- 07
Explainability & uncertainty views
Evidence note
Known training context includes patches around 100 × 100 × 60, an RTX 3060 Laptop GPU and approximately 25–50 epochs.
Result limitation
Qualitative segmentation material exists. No quantitative performance metrics are claimed here, and the uncertainty and explainability modules were not evaluated to the same extent as the main segmentation implementation.