FedDA-TSformer: Federated Domain Adaptation with vision TimeSformer for left ventricle segmentation on gated myocardial perfusion SPECT image.

Abstract

Background: Accurate assessment of left ventricular function is essential for diagnosing and managing cardiovascular disease. Gated myocardial perfusion SPECT (MPS) enables simultaneous evaluation of perfusion and function, but reliable contour extraction is challenged by image noise, resolution limits, and anatomical variability. Multi-center validation is further restricted by data privacy concerns, underscoring the need for robust and privacy-preserving contouring methods.

Methods: In this study, we propose a novel approach, FedDA-TSformer, which integrates Federated Domain Adaptation with the TimeSformer model for the task of left ventricle segmentation using MPS images. The proposed model captures spatial and temporal features through a Divide-Space-Time-Attention mechanism, which ensures spatial-temporal consistency in predictions across multi-center datasets. To facilitate domain adaptation, we employ a local maximum mean discrepancy (LMMD) loss to align model outputs across data from three different institutions. This strategy effectively combines federated learning and domain adaptation to enhance model generalization while ensuring data security.

Results: We evaluated FedDA-TSformer on a dataset comprising 150 subjects collected from three hospitals, with each cardiac cycle divided into eight gates. The model achieved Dice Similarity Coefficients (DSC) of 0.842 and 0.907 for left ventricular endocardium and epicardium segmentation, respectively.

Discussion: FedDA-TSformer provides a robust, privacy-preserving solution for multi-center left ventricular segmentation, outperforming traditional FedAvg in handling domain shifts. By leveraging the TimeSformer architecture and domain adaptation mechanisms, the framework ensures spatial-temporal consistency and data security across heterogeneous clinical sites. Despite current limitations regarding communication overhead and its focus on a small SPECT-only dataset, this study establishes a scalable foundation for collaborative cardiac diagnosis. Future work will prioritize model compression, asynchronous updates, and cross-modality generalization to CT and MRI to enhance its practicality in resource-constrained environments.