基于注意力机制的急性胰腺炎影像分割研究  

Research on image segmentation of acute pancreatitis based on attention mechanism

作　　者：邓鸿 肖佳丽 冯文 祝元仲 肖波[2,3] 何汶静[1] Deng Hong;Xiao Jiali;Feng Wen;Zhu Yuanzhong;Xiao Bo;He Wenjing(School of Medical Imaging,North Sichuan Medical College,Nanchong 637000,China;Bishan Hospital of Chongqing,Bishan hospital of Chongqing medical university,Chongqing 402760,China;Department of Radiology,the Affiliated Hospital of North Sichuan Medical College,Nanchong 637000,China)

机构地区：[1]川北医学院医学影像学院,南充637000 [2]重庆医科大学附属璧山医院(重庆市璧山区人民医院),重庆402760 [3]川北医学院附属医院放射科,南充637000

出　　处：《国际生物医学工程杂志》2024年第2期141-148,共8页International Journal of Biomedical Engineering

基　　金：南充市市校科技战略合作项目(22SXQT0323);川北医学院博士启动资金(CBY21-QD22);川北医学院重点培育项目(CBY21-ZD02)。

摘　　要：目的探究卷积注意力模块(CBAM)与Unet不同融合路径对于急性胰腺炎患者增强CT图像中胰腺自动分割的有效性。方法回顾性分析川北医学院附属医院2016年1月1日至2021年7月30日收治的1158例急性胰腺炎患者,纳入首发急性胰腺炎患者141例,依据轻、中、重病例各随机选取5例共15例作为测试集,余下126例作为训练集,在训练集中再随机划分20%的数据作为验证集。以Dice相似系数、豪斯多夫距离(HD)和像素准确率(PA)作为评价指标,对CBAM与Unet网络的不同融合路径进行训练,取验证集表现最佳的模型,在训练集上评估其性能,并将其与Unet、在跳级连接部分加入了注意门注意力机制(AttentionUnet)、在Unet网络中用ResBolck替代原有的卷积模块(ResUnet)、在特征提取的跳级连接分支模块融入CBAM(ResUnet_CBAM)模型进行比较。结果Unet_CBAM在训练集上取得的效果更好,Dice相似系数为80.06%,HD为3.7659,PA为0.9923,均优于其他融合路径,对急性胰腺炎患者CT图像中胰腺区域的分割效果均优于Unet及其相关的变体网络。结论Unet网络在跳级连接后融入CBAM能够较好地对急性胰腺炎患者增强CT图像行胰腺分割,能有效地提升相关人员对急性胰腺炎患者增强CT图像进行胰腺分割的效率。Objective To assess the efficacy of different fusion strategies involving the convolutional block attention module(CBAM)and Unet for automatic pancreas segmentation in enhanced CT images of patients with acute pancreatitis.Methods A retrospective analysis was conducted on 1158 patients with acute pancreatitis admitted to the Affiliated Hospital of North Sichuan Medical College between January 1st,2016 and July 30th,2021.Among them,141 patients with first-episode acute pancreatitis were randomly categorized into mild,moderate,and severe cases.The test set comprised 5 mild and 15 severe cases,while the remaining 126 cases were used for training.Within the training set,20%of the data was randomly allocated as the validation set.Different fusion paths of the CBAM and Unet networks were trained,utilizing the Dice similarity coefficient,Hausdorff distance(HD),and pixel accuracy(PA)as evaluation metrics.The model demonstrating the best performance on the validation set was selected and evaluated on the test set.Additionally,the Unet model was combined with the attention gate attention mechanism(AttentionUnet)in the skip connection,and the ResBlock replaced the original convolution module(ResUnet)in the Unet network.Moreover,the skip connection branch module of feature extraction was integrated with CBAM(ResUnet_CBAM)for comparison.Results Unet_CBAM achieved better results on the test set with a Dice value of 80.06%,a HD value of 3.7659 and a PA value of 0.9923,all surpassing other fusion strategies.The segmentation accuracy of the pancreatic region in CT images of acute pancreatitis patients was notably enhanced compared to Unet and its related variant networks.Conclusions The Unet network integrated into CBAM after skip connection can better perform pancreatic segmentation on enhanced CT images of patients with acute pancreatitis and can effectively improve the efficiency of relevant personnel in pancreatic segmentation on enhanced CT images of patients with acute pancreatitis.

关 键 词：急性胰腺炎 注意力机制 医学图像处理 胰腺分割 

分 类 号：R816.5[医药卫生—放射医学] R576[医药卫生—临床医学]