Wang Zhen, Fu Shuang, Zhang Hongguang, Wang Chunyang, Xia Chunhui, Hou Pen, Shun Chunxue, Shun Ge
School of Public Health, Qiqihar Medical University, Qiqihar, 161003, China.
College of Pharmacy, Qiqihar Medical University, Qiqihar, 161003, China.
Sci Rep. 2025 Mar 10;15(1):8194. doi: 10.1038/s41598-025-92715-0.
Accurate segmentation of organs or lesions from medical images is essential for accurate disease diagnosis and organ morphometrics. Previously, most researchers mainly added feature extraction modules and simply aggregated the semantic features to U-Net network to improve the segmentation accuracy of medical images. However, these improved U-Net networks ignore the semantic differences of different organs in medical images and lack the fusion of high-level semantic features and low-level semantic features, which will lead to blurred or miss boundaries between similar organs and diseased areas. To solve this problem, we propose Dual-branch dynamic hierarchical U-Net with multi-layer space fusion attention (D2HU-Net). Firstly, we propose a multi-layer spatial attention fusion module, which makes the shallow decoding path provide predictive graph supplement to the deep decoding path. Under the guidance of higher semantic features, useful context features are selected from lower semantic features to obtain deeper useful spatial information, which makes up for the semantic differences between organs in different medical images. Secondly, we propose a dynamic multi-scale layered module that enhances the multi-scale representation of the network at a finer granularity level and selectively refines single-scale features. Finally, the network provides guiding optimization for subsequent decoding based on multi-scale loss functions. The experimental results on four medical data sets show D2HU-Net enables the most advanced segmentation capabilities on different medical image datasets, which can help doctors diagnose and treat diseases.
从医学图像中准确分割器官或病变对于准确的疾病诊断和器官形态测量至关重要。以前，大多数研究人员主要是在U-Net网络中添加特征提取模块并简单地聚合语义特征，以提高医学图像的分割精度。然而，这些改进的U-Net网络忽略了医学图像中不同器官的语义差异，缺乏高级语义特征和低级语义特征的融合，这将导致相似器官和病变区域之间的边界模糊或遗漏。为了解决这个问题，我们提出了具有多层空间融合注意力的双分支动态分层U-Net（D2HU-Net）。首先，我们提出了一种多层空间注意力融合模块，该模块使浅层解码路径为深层解码路径提供预测图补充。在更高语义特征的指导下，从较低语义特征中选择有用的上下文特征，以获得更深层次的有用空间信息，这弥补了不同医学图像中器官之间的语义差异。其次，我们提出了一种动态多尺度分层模块，该模块在更精细的粒度级别上增强了网络的多尺度表示，并选择性地细化单尺度特征。最后，该网络基于多尺度损失函数为后续解码提供指导优化。在四个医学数据集上的实验结果表明，D2HU-Net在不同的医学图像数据集上具有最先进的分割能力，这可以帮助医生进行疾病诊断和治疗。
