Scholarlink

<p>The spine is essential for supporting the human body and maintaining its functions. In primary healthcare settings, radiologist shortages and medical resource scarcity often result in spinal fracture misdiagnoses and oversights. This paper introduces a spinal segmentation method utilizing heatmap regression and an enhanced UNet, aiding radiologists in swiftly diagnosing spinal fractures and enhancing diagnostic efficiency. During landmark positioning, this paper presents an M-SCN network structure designed for the fusion of multiple features. The LA stage of the SCN incorporates multi-level feature outputs, while the SC stage employs cyclic feature fusion, enhancing the feature extraction method and the selection of the maximum feature values, thereby improving the ability of SCN to capture both local and global image features. To enhance positioning accuracy, we propose a graph-structured multi-coordinate point optimization method. A weighted directed graph is constructed, employing the predicted heat map’s peak value and the Euclidean distances between points to define unary terms and pairwise weights for graph edges. Finally, the Bellman-Ford algorithm optimizes the coordinate points, significantly enhancing positioning accuracy. Compared to the original SCN network, the M-SCN network reduces the prediction error rate by 4.43%. To address UNet’s inability to capture global feature information, which leads to incomplete vertebral segmentation, we introduce a UNet-CB network structure that incorporates global context attention. ContextBlocks and residual connections are integrated into the UNet BasicBlock to enhance global feature capture. Furthermore, a central heatmap is introduced to improve the recognition of individual vertebrae. In the vertebral segmentation experiment, the Dice Similarity Coefficient (DSC) reached 95.64%, the Maximum Symmetry Surface Distance (MSSD) measured 4.862&#xa0;mm, and the accuracy and recall rates were 95.86 and 92.46%, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spinal segmentation based on heatmap regression positioning and improved UNet

  • 期刊论文

  • 发表日期：

    2025年8月19日

查看全文
  • Yuyao Huang,
  • Yuhang Wang,
  • Li He,
  • Zhiqin He,
  • Lin Xiao

摘要

The spine is essential for supporting the human body and maintaining its functions. In primary healthcare settings, radiologist shortages and medical resource scarcity often result in spinal fracture misdiagnoses and oversights. This paper introduces a spinal segmentation method utilizing heatmap regression and an enhanced UNet, aiding radiologists in swiftly diagnosing spinal fractures and enhancing diagnostic efficiency. During landmark positioning, this paper presents an M-SCN network structure designed for the fusion of multiple features. The LA stage of the SCN incorporates multi-level feature outputs, while the SC stage employs cyclic feature fusion, enhancing the feature extraction method and the selection of the maximum feature values, thereby improving the ability of SCN to capture both local and global image features. To enhance positioning accuracy, we propose a graph-structured multi-coordinate point optimization method. A weighted directed graph is constructed, employing the predicted heat map’s peak value and the Euclidean distances between points to define unary terms and pairwise weights for graph edges. Finally, the Bellman-Ford algorithm optimizes the coordinate points, significantly enhancing positioning accuracy. Compared to the original SCN network, the M-SCN network reduces the prediction error rate by 4.43%. To address UNet’s inability to capture global feature information, which leads to incomplete vertebral segmentation, we introduce a UNet-CB network structure that incorporates global context attention. ContextBlocks and residual connections are integrated into the UNet BasicBlock to enhance global feature capture. Furthermore, a central heatmap is introduced to improve the recognition of individual vertebrae. In the vertebral segmentation experiment, the Dice Similarity Coefficient (DSC) reached 95.64%, the Maximum Symmetry Surface Distance (MSSD) measured 4.862 mm, and the accuracy and recall rates were 95.86 and 92.46%, respectively.