融合双注意力机制的多尺度胰腺分割方法  

Multiscale pancreas segmentation method based on double attention mechanism

作　　者：张国栋 唐晓艺[1] 鞠蓉晖 宫照煊 ZHANG Guo-dong;TANG Xiao-yi;JU Rong-hui;GONG Zhao-xuan(School of Computer Science,Shenyang Aerospace University,Shenyang 110136,China;Key Laboratory of Intelligent Computing in Medical Image of Ministry of Education,Northeastern University,Shenyang 110819,China;Radiology Department,The People’s Hospital of Liaoning Province,Shenyang 110067,China)

机构地区：[1]沈阳航空航天大学计算机学院,辽宁沈阳110136 [2]东北大学医学影像智能计算教育部重点实验室,辽宁沈阳110819 [3]辽宁省人民医院放射科,辽宁沈阳110067

出　　处：《计算机工程与设计》2024年第4期1189-1194,共6页Computer Engineering and Design

基　　金：辽宁省自然科学基金项目(2020-MS-239、2019-ZD-0234);辽宁省教育厅基金项目(LJKZ0210、JYT19053、JYT19040);航空科学基金项目(2019ZE054009)。

摘　　要：为解决CT图像中胰腺边界不规则导致分割精度不高的问题,提出一种融合双注意机制的多尺度U型网络模型。该模型由一个编码器及两个解码器组成,提高特征利用。针对模型中连续下采样导致特征空间信息损失的问题,提出一种金字塔注意力特征融合模块,引入通道和空间两个独立注意力机制,提供多尺度输入信息并行采样,提高边界提取性能,提升分割精度。实验结果表明,该方法在ISICDM 2018数据集上的平均Dice系数为85.35%,具有效性。Aiming at the problem of low segmentation accuracy caused by the irregular boundary of pancreas in CT image,a dilated space pyramid pooling U network integrating double attention mechanism was proposed.The model was composed of an encoder and two decoders to improve feature utilization.Aiming at the problem that continuous down sampling operation may lead to the loss of feature space information,a pyramid attention feature fusion module for parallel sampling was proposed.The channel attention mechanism and the space attention mechanism were constructed.The multi-scale input information for context feature extraction was provided,which increased the receptive field and further improved the segmentation accuracy.Experimental results show that the average Dice coefficient of the proposed method is 85.35%for the ISICDM 2018 dataset.

关 键 词：胰腺分割 注意力机制 双解码器 金字塔池化 特征融合 边界提取 多尺度信息 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]