ECHO-SegNet Framework for Accurate 2D Echocardiographic Image Segmentation Using the CAMUS Dataset

Accurate segmentation of cardiac structures in echocardiographic images plays a vital role in the quantitative assessment of cardiac function, enabling early detection and management of cardiovascular diseases. This study presents ECHO-SegNet, a deep learning-based framework designed for automated segmentation of the left ventricle (LV), myocardium, and left atrium (LA) in 2D echocardiographic images using the publicly available CAMUS dataset, which includes apical four-chamber (A4C) and two-chamber (A2C) views from 500 patients with expert annotations. The designed model combines an efficient U-Net backbone network with attention and deep supervision to improve boundary accuracy and segmentation consistency, even for difficult low-contrast or artifact-contaminated ultrasound images. The innovation in this work is the introduction of a multi-scale context-aware attention module and a temporal frame consistency regularizer, which collectively allow the model to adaptively pay attention to useful cardiac structures and preserve spatial consistency between successive frames. Large-scale testing on the CAMUS dataset proves the success of the introduced model with a Dice Similarity Coefficient (DSC) of 0.9157, accuracy of 96.83% and a DSC on validation of 0.8950 and a validation accuracy of 96.83%. The model achieved a low validation loss of 0.0994 representing consistent and accurate performance across varying patient cases. These findings confirm that ECHO-SegNet exceeds baseline models with a new state-of-the-art in 2D echocardiographic segmentation for real-time clinical use.