This repository contains the implementation of the MICCAI 2025 early accepted paper:

RSAD: Region-Specific Anomaly Detection in fMRI for Disease Diagnosis

RSAD is a novel framework for detecting functional anomalies in brain fMRI data, focusing on region-level discrepancy modeling to improve disease diagnosis. It leverages attention mechanisms and region-wise reconstruction losses to enhance sensitivity to subtle neural abnormalities.

• ✅ We propose the RSAD framework, which adapts fMRI-based disease diagnosis to anomaly detection tasks and achieves high performance under extreme class imbalance conditions.
• 🎯 We design an affinity matrix learning module and an adaptive Region of Interest (ROI) masking strategy to capture relationships between brain regions, significantly enhancing the representational power of masked autoencoders for fMRI signals.
• 🧮 We introduce a region-specific discrepancy score weighting strategy that amplifies the distinctions between potential anomalies and normal controls, effectively improving the anomaly detection performance of pre-trained models.
• 📈 Our approach is evaluated on six distinct brain disease datasets, and the experimental results demonstrate its superiority over existing state-of-the-art methods.

git clone https://github.com/kylin1112/RSAD.git
cd RSAD

pip install -r requirements.txt

Use toolkit/data_preprocess_tools.py to preprocess your own fMRI data in standard space. This tool supports ROI extraction based on different atlases and allows splitting into training, validation, and test sets.

Run train.py to perform self-supervised pretraining of the RSAD model.

python train.py

@inproceedings{sun2025rsad,
  title     = {RSAD: Region-Specific Anomaly Detection in fMRI for Disease Diagnosis},
  author    = {Yusong Sun and Dongdong Chen and Mengjun Liu and Zhenrong Shen and Zhiyun Song and Yuqi Hu and Manman Fei and Xu Han and Zelin Liu and Xingkai Fang and Lu Bai and Lichi Zhang},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
  year      = {2025}
}