Overview
This project is an implementation of the MICCAI 2025 paper: “Inter-class separability loss for weakly supervised mutually exclusive multiclass segmentation of brain tumor lesions” by Vivek Dhamale and Vaanathi Sundaresan.
Subset Run Analysis
Due to computational constraints, the approach was tested using a subset of 6 subjects (4 training, 1 validation, 1 test) from the BraTS 2020 dataset, instead of the full 237 training subjects.
Key Results
| Metric | Subset Run (6 Subjects) | Full Paper Results (237 Subjects) |
|---|---|---|
| Tumor Core Dice | 0.665 | 0.766 |
| Tumor Core IoU | 0.665 | - |
| Tumor Core HD95 | 125.18 mm | 55.09 mm |
| Edema Dice | 0.730 | 0.745 |
Observations
- Despite using only 6 subjects (representing just ~1.6% of the original training data), the obtained Edema Dice score of 0.730 is remarkably close to the authors’ 0.745 using 237 subjects. This demonstrates that the proposed Binary-Guided Structural Prior requires minimal data to successfully learn and localize macroscopic edema structure without any pixel-level supervision. This efficiency is largely attributed to the contrastive pretraining.
- The low performance on the tumor core indicates that 6 subjects is insufficient to learn detailed, intricate tumor structures.