TimeSformer-Based Federated Domain Adaptation for Multi-Site Automatic Left Ventricular Segmentation and Quantification on Gated Myocardial Perfusion SPECT Images
Disciplines
Bioimaging and Biomedical Optics
Abstract (300 words maximum)
Cardiovascular disease (CVD) is a leading cause of morbidity and mortality worldwide, responsible for millions of deaths annually and placing immense pressure on healthcare systems. Early diagnosis and effective management of conditions like coronary artery disease (CAD) are crucial for preventing severe outcomes such as heart failure and myocardial infarction. One key diagnostic tool for CAD is gated myocardial perfusion single-photon emission computed tomography (MPS), which provides detailed, phase-specific images of the heart throughout the cardiac cycle, making it essential for assessing left ventricular (LV) function. Despite its effectiveness, multi-center studies using MPS face challenges due to the manual segmentation of LV contours, which is labor-intensive and time-consuming. Additionally, sharing patient data between institutions raises privacy concerns. To address these issues, this study proposes a deep learning-based algorithm for automated LV contour extraction from MPS images, with a focus on data privacy through federated learning. Our method utilizes a TimeSformer model combined with FedDAvT. The modified TimeSformer model processes 3D volumetric data and learns temporal sequences within the volumes, capturing temporal correlations between cardiac phases. FedDAvT, a federated learning approach, offers superior domain adaptation and ensures patient privacy, as no raw data is shared between institutions. Our model was trained using 150 fully deidentified MPS datasets collected from three hospitals: 73 from the First Affiliated Hospital of Nanjing Medical University, 28 from Chang Bing Show Chwan Memorial Hospital, Taiwan, and 49 from Xiangya Hospital, Central South University. Our FedDA-TSformer model achieved a Dice Similarity Coefficient (DSC) of 0.842 for the endocardium and 0.907 for the epicardium in left ventricle segmentation, demonstrating the effectiveness of our model in accurately segmenting the left ventricle.
Academic department under which the project should be listed
CCSE - Computer Science
Primary Investigator (PI) Name
Chen Zhao
TimeSformer-Based Federated Domain Adaptation for Multi-Site Automatic Left Ventricular Segmentation and Quantification on Gated Myocardial Perfusion SPECT Images
Cardiovascular disease (CVD) is a leading cause of morbidity and mortality worldwide, responsible for millions of deaths annually and placing immense pressure on healthcare systems. Early diagnosis and effective management of conditions like coronary artery disease (CAD) are crucial for preventing severe outcomes such as heart failure and myocardial infarction. One key diagnostic tool for CAD is gated myocardial perfusion single-photon emission computed tomography (MPS), which provides detailed, phase-specific images of the heart throughout the cardiac cycle, making it essential for assessing left ventricular (LV) function. Despite its effectiveness, multi-center studies using MPS face challenges due to the manual segmentation of LV contours, which is labor-intensive and time-consuming. Additionally, sharing patient data between institutions raises privacy concerns. To address these issues, this study proposes a deep learning-based algorithm for automated LV contour extraction from MPS images, with a focus on data privacy through federated learning. Our method utilizes a TimeSformer model combined with FedDAvT. The modified TimeSformer model processes 3D volumetric data and learns temporal sequences within the volumes, capturing temporal correlations between cardiac phases. FedDAvT, a federated learning approach, offers superior domain adaptation and ensures patient privacy, as no raw data is shared between institutions. Our model was trained using 150 fully deidentified MPS datasets collected from three hospitals: 73 from the First Affiliated Hospital of Nanjing Medical University, 28 from Chang Bing Show Chwan Memorial Hospital, Taiwan, and 49 from Xiangya Hospital, Central South University. Our FedDA-TSformer model achieved a Dice Similarity Coefficient (DSC) of 0.842 for the endocardium and 0.907 for the epicardium in left ventricle segmentation, demonstrating the effectiveness of our model in accurately segmenting the left ventricle.