Performance Comparison of Deep Learning Models for Multi-Structure Cardiac Segmentation in Two-Chamber Echocardiography Images
Abstract
Two-dimensional echocardiography is a widely adopted non-invasive medical imaging modality for the assessment of cardiac structure and function. However, manual delineation of cardiac boundaries remains a time-consuming task requiring substantial clinical expertise. This study presents a comparative evaluation of three deep learning architectures Fully Convolutional Network (FCN), U-Net, and DeepLabV3+ for automated multi-structure cardiac segmentation in Apical Two-Chamber echocardiographic views using the publicly available CAMUS dataset. Model performance was assessed via $\mathbf{1 0}$-fold cross-validation, with the Dice Similarity Coefficient (DSC) serving as the primary evaluation metric. Segmentation outputs from the best-performing model were further utilized to estimate Left Ventricular Ejection Fraction (LVEF) for cardiac function assessment. Experimental results demonstrate that DeepLabV3+ achieves superior overall performance, attaining mean DSC values of 0.938 and 0.941 for the End-Systole (ES) and End-Diastole (ED) phases, respectively, outperforming U-Net (0.893, 0.894) and FCN (0.871, 0.888). The performance advantage of DeepLabV3+ is attributed to its Atrous Spatial Pyramid Pooling mechanism, which enables effective multi-scale feature extraction and enhances delineation accuracy across heterogeneous cardiac structures. These findings suggest that DeepLabV3+ holds strong potential for deployment in automated echocardiographic analysis pipelines to support clinical decision-making in cardiac disease diagnosis.