Back to Projects
Brain Tumor Segmentation
Investigated 3D deep-learning architectures for brain tumor segmentation using multimodal MRI data from the BraTS dataset.
PythonPyTorch3D U-NetAttention U-NetEfficientNetMONAI
Related Projects
SYSTEM ARCHITECTURE
Best validation Dice: 0.8256 using the Attention + EfficientNet architecture.
Problem
Volumetric tumor segmentation requires models to identify relevant structures across three-dimensional MRI scans while managing substantial memory demands.
Architecture
MRI preprocessing → volumetric patch sampling → 3D encoder-decoder segmentation models → Dice and IoU evaluation.
Challenges
- Prepared and normalized volumetric MRI data
- Compared baseline and attention-enhanced 3D segmentation architectures
- Managed GPU memory constraints during patch-based training
- Evaluated segmentation quality using Dice score and IoU
Benchmarks
The project explored baseline 3D U-Net, EfficientNet-encoder, and attention-enhanced variants. Include exact metric values only when they match the final reproducible validation run.
Lessons Learned
- Three-dimensional models require careful memory and patch-size management
- Architecture comparisons should use consistent validation methodology
- Dice and IoU provide complementary views of segmentation quality