Fed-CSA: Channel Spatial Attention and Adaptive Weights Aggregation-Based Federated Learning for Breast Tumor Segmentation on MRI.

Breast tumor segmentation is an important task in medical image analysis, especially in the diagnosis and treatment of breast cancer. Magnetic resonance imaging (MRI) is a widely used imaging modality for breast tumor segmentation due to its high spatial and temporal resolution. However, there are several challenges associated with breast tumor segmentation on MRI data. One of the main challenges is the heterogeneity of MRI data collected from different imaging centers, which is usually caused by inconsistencies in the imaging protocols, quality, and resolution. Additionally, there are privacy concerns associated with the use of patient data, which limit the sharing of data across institutions. To address these challenges, we propose a novel federated learning framework called Federated Channel Spatial attention adaptive weight Aggregation (Fed-CSA) for breast tumor segmentation on MRI data. The proposed framework uses a CSA module to handle the heterogeneity of MRI data collected from different centers, which can pay more attention to the target regions and ignore non-related regions. Furthermore, the adaptive weight aggregation (AWA) method is proposed to improve the efficiency of model weight aggregation in federated learning across multiple clients while preserving patients’ privacy. The proposed Fed-CSA framework is trained and tested on a large-scale dataset of 1460 patients from different centers, and the results demonstrate its effectiveness in terms of efficiency and feasibility, outperforming other baseline algorithms.