Cao Kerang, Zhao Miao, Geng Minghui, Zheng Shuai, Jung Hoekyung
College of Computer Science and Technology, Shenyang University of Chemical Technology, Shenyang, China.
Key Laboratory of Intelligent Technology of Chemical Process Industry in Liaoning Province, Shenyang, China.
PLoS One. 2025 Jun 25;20(6):e0325794. doi: 10.1371/journal.pone.0325794. eCollection 2025.
This paper introduces an optimized nested UNet model for automated left ventricular segmentation in cardiac function assessment. We utilize the EchoNet-Dynamic dataset, which contains both video data and expert annotations. Unlike conventional methods such as DeepLabv3 that struggle with large model sizes and imprecise segmentation, Our proposed model introduces a deeper feature extraction module to effectively capture multi-scale features and reduce computational overhead. By integrating the CBAM (Attention module) attention mechanism and a lightweight SimAM (Simple Attention Module) module, we enhance feature selectivity and minimize redundancy. To further stabilize training and address gradient issues, we combine binary cross-entropy and Dice loss functions. Experimental results reveal that our model significantly outperforms existing methods, achieving a 1.05% increase in the Dice coefficient and reducing model size to 15% of the original. These improvements not only enhance the accuracy of cardiac function assessments but also provide a more efficient solution for automated diagnosis in clinical practice.
本文介绍了一种用于心功能评估中自动左心室分割的优化嵌套UNet模型。我们使用了包含视频数据和专家注释的EchoNet-Dynamic数据集。与诸如DeepLabv3等存在模型规模大及分割不精确问题的传统方法不同，我们提出的模型引入了一个更深的特征提取模块，以有效捕获多尺度特征并减少计算开销。通过集成CBAM（注意力模块）注意力机制和轻量级SimAM（简单注意力模块）模块，我们提高了特征选择性并最小化冗余。为了进一步稳定训练并解决梯度问题，我们结合了二元交叉熵和Dice损失函数。实验结果表明，我们的模型显著优于现有方法，Dice系数提高了1.05%，模型规模减小到原来的15%。这些改进不仅提高了心功能评估的准确性，还为临床实践中的自动诊断提供了更高效的解决方案。
