Key result
Residual 3D-UNet segments left atrium with a ~0.94 Dice score on dynamic CT.
Why the study?
Automatic segmentation of the left atrium across all cardiac phases remains challenging due to substantial anatomical and image variations throughout the cardiac cycle.
Does an AI-based framework utilizing a Residual 3D-UNet and OCSVM accurately segment the left atrium from ECG-gated CT scans and reliably classify atrial fibrillation versus sinus rhythm?
Does an AI-based framework utilizing a Residual 3D-UNet and OCSVM accurately segment the left atrium from ECG-gated CT scans and reliably classify atrial fibrillation versus sinus rhythm?
A comprehensive deep learning framework combining a Residual 3D-UNet and OCSVM successfully automates left atrial segmentation across the cardiac cycle and reliably classifies rhythm status, offering a potential tool to enhance atrial fibrillation management.
Atrial fibrillation (AF) is a heart condition widely recognized as a significant risk factor for stroke. Left atrial (LA) volume variation has been identified as a key predictor of AF, and several researchers have proposed deep learning models capable of quickly providing this measurement by processing computed tomography (CT) or magnetic resonance images. In clinical imaging, time-varying ECG-gated CT offers precise information about LA anatomy and function, which could help in developing personalized treatment plans for AF patients. Furthermore, advancements in time-varying dataset acquisition indicate the potential for expanding the role of CT in the management of AF patients through specialized processing techniques. However, automatic segmentation of the LA across all cardiac phases remains challenging due to significant variations in both anatomical structures and image signals throughout the cardiac cycle. To overcome these challenges, this study presents a comprehensive AI-based framework designed to segment the LA across the entire cardiac cycle and classify patients with AF. Specifically, our framework employs a customized Residual 3D-UNet model to segment the LA from time-varying ECG-gated CT scans and utilizes a One-Class Support Vector Machine (OCSVM) to distinguish patients in sinus rhythm (SR) from those with AF. A dataset of 93 time-varying ECG-gated CT scans was retrospectively collected: 60 patients were used for the segmentation task, while 33 patients were used for the classification task. The Residual 3D-UNet model demonstrated high accuracy, achieving a mean Dice score of 0.94, with consistent precision (94.45%) and recall (94.83%) across ten cardiac phases. The OCSVM achieved 78.7% accuracy with high specificity (86.3%), effectively minimizing the risk of misclassifying AF as SR, although sensitivity was lower at 70%, demonstrating the potential of automated segmentation and rhythm classification, providing a potential valuable tool for AF diagnosis.
No takes yet. Share an insight, caveat, or question.
Buongiorno et al. (2025) studied Non-valvular atrial fibrillation (n=93). Residual 3D-UNet and One-Class Support Vector Machine (OCSVM) vs. Manual segmentation and clinical rhythm evaluation was evaluated on Mean Dice score for left atrial segmentation across the entire cardiac cycle. The customized Residual 3D-UNet model achieved a mean Dice score of 0.94 for left atrial segmentation, and the OCSVM model classified sinus rhythm versus atrial fibrillation with 78.7% accuracy.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: