Hypergraph Tversky-Aware Domain Incremental Learning for Brain Tumor Segmentation with Missing Modalities

• ReHyDIL — Replay-based Hypergraph Domain Incremental Learning

• MICCAI 2025

• Paper link

Dataset Preparation

This work requires the BraTS 2019 (BraTS19) dataset. You can request access and download it from the official source:

• The Brain Tumor Segmentation (BraTS) Challenge

The script expects the data to be in a directory named ./BraTS19

./BraTS19/
├── HGG/
│   ├── BraTS19_TCIA01_.../
│   │   ├── BraTS19_TCIA01_..._flair.nii.gz
│   │   ├── BraTS19_TCIA01_..._t1.nii.gz
│   │   ├── BraTS19_TCIA01_..._t1ce.nii.gz
│   │   ├── BraTS19_TCIA01_..._t2.nii.gz
│   │   └── BraTS19_TCIA01_..._seg.nii.gz
│   └── ...
└── LGG/
    ├── BraTS19_TCIA08_.../
    │   ├── ...
    └── ...

Preprocess

pre.py is used to convert all .nii/.nii.gz volumes to .npy format and perform the full preprocessing pipeline (e.g., orientation/spacing standardization, normalization, cropping/padding, and split generation).

python pre.py

Create patient-level train/val/test lists using a 80/10/10 split (patient-level) from BraTS19. Adjust paths/ratios as needed.

python pre_list.py \
  --data_root ./BraTS19 \
  --out_dir   ./lists \
  --val_ratio 0.10 \
  --test_ratio 0.10 

Train

train.py is a stage-wise runner: it trains the model incrementally over MRI modalities. Default configuration follows the clinical order: t1, t2, flair, t1ce.

If you don’t pass --stages, the script will run all four stages in that order.

Quick start (full clinical sequence — default) The model learns its first task using only the T1 modality.

python train.py \
  --data_path /path/to/data_root \
  --out_root  /path/to/outputs \
  --train_fmt /path/lists/train.list \
  --val_fmt   /path/lists/val.list

Citation

@inproceedings{wang2025hypergraph,
  title={Hypergraph tversky-aware domain incremental learning for brain tumor segmentation with missing modalities},
  author={Wang, Junze and Fan, Lei and Jing, Weipeng and Di, Donglin and Song, Yang and Liu, Sidong and Cong, Cong},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2025},
  pages={283--293},
  year={2025},
  organization={Springer}
}