<p>Echocardiographic images play a pivotal role in the diagnosis and evaluation of various cardiovascular diseases, including coronary heart disease, heart failure, and congenital heart disease. Accurate segmentation of core structures such as the left ventricle (LV) endocardium, LV epicardium, and left atrium (LA) from these images is critical for quantifying key cardiac function parameters (e.g., ejection fraction, ventricular wall thickness), which serves as a fundamental basis for clinicians to make diagnostic judgments and develop treatment plans. However, automatic segmentation of regions of interest in echocardiographic images remains a challenging task. These images exhibit inherent drawbacks like poor contrast, missing edges, low signal-to-noise ratio, and speckle noise, which increase the difficulty of defining boundaries between the LV and LA. Furthermore, existing convolutional neural networks fail to effectively utilize extracted semantic and spatial information, resulting in subpar segmentation performance. To address these issues, we propose a two-branch multi-objective semantic segmentation network named SemenNet, which is based on enhancing semantic and spatial information. Specifically, a semantic information enhancement process is designed to better learn the semantic contextual correlation between a pixel and its neighbors, and use this correlation to correct the original feature maps. In addition, a spatial enhancement process is proposed to effectively process feature maps containing rich spatial structural information. After effective fusion of the two types of feature maps, they are upsampled for segmentation. Extensive experiments conducted on the CAMUS dataset yield impressive results: the average accuracy reaches 98.12%, with Dice coefficients of 92.15% for LV endocardium, 86.95% for LV epicardium, and 87.01% for LA. These results demonstrate that SemenNet can achieve superior performance compared to other methods without relying on a heavy backbone. For medical doctors and medical staff, this work provides an efficient and accurate automatic segmentation tool: it reduces the tedious manual segmentation workload, minimizes subjective errors caused by manual operation, and quickly provides precise structural segmentation results to support the quantitative assessment of cardiac function, thereby assisting in more timely and accurate clinical diagnosis and treatment decision-making. The code for this work is available at: <a href="https://github.com/Stargazer2000/SemenNet-Echo.git.">https://github.com/Stargazer2000/SemenNet-Echo.git.</a></p>

A two-branch information-enhanced multi-objective semantic segmentation network for echocardiographic images

期刊论文
发表日期：

2025年12月26日

查看全文

Yi Ding,
Yilin Zhang,
Xiaoyong Zhang,
Mingsheng Cao,
Xianmin Wang

摘要

Echocardiographic images play a pivotal role in the diagnosis and evaluation of various cardiovascular diseases, including coronary heart disease, heart failure, and congenital heart disease. Accurate segmentation of core structures such as the left ventricle (LV) endocardium, LV epicardium, and left atrium (LA) from these images is critical for quantifying key cardiac function parameters (e.g., ejection fraction, ventricular wall thickness), which serves as a fundamental basis for clinicians to make diagnostic judgments and develop treatment plans. However, automatic segmentation of regions of interest in echocardiographic images remains a challenging task. These images exhibit inherent drawbacks like poor contrast, missing edges, low signal-to-noise ratio, and speckle noise, which increase the difficulty of defining boundaries between the LV and LA. Furthermore, existing convolutional neural networks fail to effectively utilize extracted semantic and spatial information, resulting in subpar segmentation performance. To address these issues, we propose a two-branch multi-objective semantic segmentation network named SemenNet, which is based on enhancing semantic and spatial information. Specifically, a semantic information enhancement process is designed to better learn the semantic contextual correlation between a pixel and its neighbors, and use this correlation to correct the original feature maps. In addition, a spatial enhancement process is proposed to effectively process feature maps containing rich spatial structural information. After effective fusion of the two types of feature maps, they are upsampled for segmentation. Extensive experiments conducted on the CAMUS dataset yield impressive results: the average accuracy reaches 98.12%, with Dice coefficients of 92.15% for LV endocardium, 86.95% for LV epicardium, and 87.01% for LA. These results demonstrate that SemenNet can achieve superior performance compared to other methods without relying on a heavy backbone. For medical doctors and medical staff, this work provides an efficient and accurate automatic segmentation tool: it reduces the tedious manual segmentation workload, minimizes subjective errors caused by manual operation, and quickly provides precise structural segmentation results to support the quantitative assessment of cardiac function, thereby assisting in more timely and accurate clinical diagnosis and treatment decision-making. The code for this work is available at: https://github.com/Stargazer2000/SemenNet-Echo.git.