UR-002 FedDA-TSformer: Federated Domain Adaptation with Vision TimeSformer for Left Ventricle Segmentation on Gated Myocardial Perfusion SPECT Image

This study presents FedDA-TSformer, an approach for accurate left ventricle segmentation in gated myocardial perfusion single-photon emission computed tomography (MPS) images, designed to ensure both high segmentation quality and patient data privacy. By integrating federated learning with domain adaptation techniques, the proposed model leverages a novel Divide-Space-Time-Attention mechanism that effectively captures spatio-temporal correlations inherent in multi-centered MPS datasets. Domain discrepancies among data from three different hospitals are mitigated using a local maximum mean discrepancy (LMMD) loss, enabling robust performance across various clinical settings. Evaluated on a dataset comprising 150 subjects with eight distinct cardiac cycle phases, FedDA-TSformer achieved Dice Similarity Coefficients of 0.842 and 0.907 for the segmentation of the left ventricular endocardium and epicardium, respectively. These results demonstrate the model's potential to improve the functional assessment of the left ventricle while upholding stringent data security standards.