MMSeg: A Masked Autoencoder-based Multi-center Collaborative Framework for Medical Image Segmentation

The powerful performance of deep learning typically relies on extensive annotated data. However, in the field of medical image analysis, the volume of data available within individual medical institutions is often limited, and direct data sharing among centers could pose privacy issues. To address this challenge, we propose a novel collaborative framework for medical image segmentation, which enables learning from datasets across multiple centers to develop more robust and generalizable models. Specifically, the framework leverages an MAE-based encoder to extract ciphertext features, preventing the leakage of raw data. During collaborative training, we also propose a source-awared adaptive module to adapt to the distribution differences of multiple source datasets and bridge the gap between ciphertext features and original images. The effectiveness of our method is demonstrated with superior performance over other collaborative methods on two medical image segmentation tasks.