Divabhattacharya/neuroafrica-brats-glioma-segmentation

BraTS-Africa 3D glioma segmentation pipeline with MONAI, nnU-Net/SwinUNETR, Dice/NSD/HD95 tracking. # NeuroAfrica-3D: BraTS-Africa Glioma Segmentation This project packages a BraTS-style 3D MRI segmentation pipeline for tumor sub-regions: whole tumor (WT), tumor core (TC), and enhancing tumor (ET). ## Highlights - Converts BraTS case folders into nnU-Net v2 raw dataset format. - Preserves the official label convention: `1=NCR`, `2=ED`, `3=ET`. - Uses region-mode labels for WT, TC, and ET. - Supports nnU-Net full-resolution training and SwinUNETR trainer variants. - Applies connected-component cleanup for small false-positive regions. - Includes weighted probability ensembling across model variants. ## Artifacts - `neuroafrica_brats_task5_nnunet_pipeline.ipynb` - `scripts/prepare_nnunet_dataset.py` - `scripts/postprocess_predictions.py` - `scripts/weighted_probability_ensemble.py` - `trainers/nnUNetTrainer_SwinUNETR_Compat.py` - `trainers/nnUNetTrainerBN_BS5_RBT_DS_BD_PS_Compat.py` ## Metrics To Fill After Validation - Dice for WT / TC / ET - Normalized Surface Distance for WT / TC / ET - HD95 for boundary robustness - Sensitivity, precision, specificity - Inference time per case ## AWS / SageMaker Deployment Path - Store NIfTI manifests, fold splits, and predictions in S3. - Package preprocessing and inference as ECR containers. - Run fold training through SageMaker training jobs. - Run validation-mask generation through SageMaker Batch Transform. - Track Dice, NSD, HD95, sensitivity, precision, specificity, and latency in SageMaker Experiments / CloudWatch logs.