This project implements and compares deep learning models for automated brain tumor segmentation on the BraTS-2024 dataset.
We explore CNN-based (nnUNet, SegResNet) and Transformer-based (SwinUNETR) architectures to segment different tumor sub-regions from MRI scans.

• nnUNet

  • Self-configuring UNet-based architecture.
  • Supports 2D and 3D convolutions.
  • Features: deep supervision, ensembling, extensive data augmentation.
  • Hyperparameter details stored in nnUNet_hyperparameters.

• SegResNet

  • Residual network tailored for medical image segmentation.
  • Efficient in capturing hierarchical spatial features.
  • Implemented in SegResNet_Train.ipynb.

• SwinUNETR

  • Transformer-based model using shifted windows for global context.
  • Strong performance on volumetric medical imaging.
  • Implemented in swinunetr_train.py.

The models were trained to predict the following labels:

• WT – Whole Tumor
• TC – Tumor Core
• ET – Enhancing Tumor
• RC – Resection Cavity

• Training Notebooks & Scripts
  • nnUNet_train.ipynb – Training pipeline for nnUNet.
  • SegResNet_Train.ipynb – Training pipeline for SegResNet.
  • swinunetr_train.py – Training pipeline for SwinUNETR.
• Results & Visualization
  • nnUNet_results.ipynb – Qualitative and quantitative results.
  • Validation/ – Fold-wise loss/accuracy curves (progress_fold0.png → progress_fold4.png).
• Configurations
  • nnUNet_hyperparameters/ – Hyperparameter details for nnUNet.
  • dataset.json, dataset_fingerprint.json, plans.json – Dataset and preprocessing configs.

• Training and validation tracked across 5 folds with accuracy and loss curves.
• Results visualized in notebooks (nnUNet_results.ipynb) and progress plots:
  • progress_fold0.png → progress_fold4.png.

• Train and evaluate multiple segmentation models on BraTS-2024.
• Compare CNN vs Transformer-based approaches.
• Use cross-validation to ensure robust evaluation.
• Segment clinically relevant tumor regions (WT, TC, ET, RC) for potential use in diagnosis and treatment planning.

Brain tumor segmentation on BraTS-2024 dataset using nnUNet, SegResNet, and SwinUNETR. Models trained and validated on MRI scans to segment WT, TC, ET, and RC with fold-wise performance tracking and visualizations.

Deep Learning Medical Imaging Brain Tumor Segmentation MRI BraTS-2024 nnUNet SegResNet SwinUNETR Computer Vision PyTorch