- Tagline: Deep learning for clinical MRI analysis
- Category: Featured Research
- GitHub: BraTs Challenge
Context
High-stakes multimodal MRI segmentation with class imbalance and noisy labels.
Role
Lead Researcher responsible for preprocessing, architecture, training strategy, and evaluation.
Work completed
- Preprocessed T1/T2/FLAIR using TorchIO normalization, resampling, and augmentation
- Implemented 3D U-Net with attention gates and deep supervision
- Optimized with Dice + Focal loss
- Trained on A100 with AMP and validation-based early stopping
Results
- Dice > 0.85 for whole tumor
- Dice > 0.78 for tumor core
- 15% reduction in false negatives
- ~8% improvement over baseline