Enhancing breast magnetic resonance imaging segmentation with a federated semi-supervised approach

Deep learning technique boosts automatic segmentation on breast Magnetic Resonance Imaging (MRI) and assists diagnosis. However, training such deep learning models requires substantial annotated data, which is resource-intensive. This is particularly difficult to achieve for a single medical institution, especially some small ones. To this end, we proposed a federated semi-supervised learning framework for automated segmentation on breast MRI images, to maximize resource utilization of individual and collaborative institutions while preserving privacy. Specifically, each participant trains its own deep learning model locally and uploads it to the server for aggregation. For each participant's local training, we perform multiple perturbations on the unannotated samples and the extracted features for enhancing model robustness and generalization, and design a loss function to jointly optimize the model's performance on both annotated and unannotated data. Our experiments across three hospitals demonstrated that our proposed method achieved superior performance, with a Dice Similarity Coefficient (DSC) of 94.8% and an Intersection over Union (IoU) of 86.6%, outperforming existing models by up to 8.8% and 12% respectively in breast MRI segmentation.