融合Vnet和边缘特征的肺结节分割算法-《中国医学物理学杂志》

文章信息/Info

Title:: Lung nodule segmentation algorithm integrating Vnet and boundary features

作者:: 江悦莹1; 施一萍1; 翁晓俊2; 朱亚梅1; 邓源1; 刘瑾1; 1.上海工程技术大学电子电气工程学院，上海 201620； 2.高博医疗集团上海阿特蒙医院医疗质量中心，上海 200003

Author(s):: JIANG Yueying1; SHI Yiping1; WENG Xiaojun2; ZHU Yamei1; DENG Yuan1; LIU Jin1; 1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China 2. Gao Broad Healthcare Group Shanghai Artemed Hospital, Shanghai 200003, China

Keywords:: Keywords: lung nodule segmentation Vnet network dilated convolution attention mechanism boundary key point selection algorithm

摘要:: 目的：针对肺结节分割过程中存在边缘信息丢失、边界分割模糊问题，在Vnet基础上改进BSR-Vnet算法。方法：首先，利用边缘关键点选择算法和空洞卷积改进BFB边缘特征增强模块，集成进编码器中，用以保留更多边缘信息；其次，用空间-通道压缩激励模块sc-SE引入双注意力机制，代替Vnet的瓶颈结构，提取非局部上下文信息；最后，利用Vnet原有的残差结构，使用混合空洞卷积构造残差空洞模块RDB代替原本的解码器，用于扩大感受野，提取更多特征细节。结果：本文改进的BSR-Vnet应用在公共数据集LIDC-IDRI上，在Dice系数上达到了87.94%的结果。结论：与基础Vnet方法相比，该模型有效保留了更多的边缘结构和信息，提取了更多上下文信息，使肺结节分割结果更为精确。【关键词】肺结节分割；Vnet网络；空洞卷积；注意力机制；边缘关键点选择算法

Abstract:: Abstract: Objective To improve BSR-Vnet algorithm based on Vnet for solving the problems of boundary information loss and boundary segmentation blurring in the process of lung nodule segmentation. Methods The boundary feature enhancement block was constructed using boundary key point selection algorithm and dilated convolution, and then it was integrated into the encoder to retain more boundary information. Then, the double attention mechanism was introduced by spatial and channel squeeze and excitation block to replace the bottleneck structure of Vnet for extracting non-local contextual information. Finally, the original residual structure of Vnet was used to construct the residual dilated block instead of the original decoder by the hybrid null convolution, thereby expanding the perceptual field and extracting more feature details. Results The improved BSR-Vnet realized a Dice coefficient of 87.94% on the public data set LDC-IDRI. Conclusion The proposed model effectively retains more boundary structure and information, and extracts more context information, which makes the lung nodule segmentation more accurate.