融合双解码和全局注意力上采样模块的胰腺囊性肿瘤分割网络  

Pancreatic Cystic Neoplasms Segmentation Network Combining Dual Decoding and Global Attention Upsampling Modules

作　　者：戴俊龙 何聪 武杰[1] 边云 Dai Junlong;He Cong;Wu Jie;Bian Yun(School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Radiology,The First Affiliated Hospital of Naval Medical University,Shanghai 200434,China)

机构地区：[1]上海理工大学健康科学与工程学院,上海200093 [2]海军军医大学第一附属医院放射诊疗科,上海200434

出　　处：《波谱学杂志》2024年第2期151-161,共11页Chinese Journal of Magnetic Resonance

摘　　要：胰腺因其解剖结构复杂多变、周围环境复杂等特点,始终是医学图像分割中最具挑战性的任务之一.针对以上问题,提出一种融合双解码和全局注意力上采样模块的深度学习分割模型(Combining Dual Decoding and Global Attention Upsampling Modules Network,DGANet).模型由一个编码器和两个解码器构成,两个解码器实现了对不同深度特征信息的充分利用;模型采用全局注意力上采样模块(Global Attention Upsampling,GAU),利用高层丰富的语义信息来引导低层选择更为精准的特征信息.利用长海医院提供的数据集进行实验,结果表明平均Dice相似系数为86.28%,交并比(Intersection-over-Union,IoU)为0.77,豪斯多夫距离(Hausdorff Distance,HD)为7.7 mm,数据证实了该模型在胰腺囊性肿瘤分割中具有一定的临床意义和价值.The pancreas has always been one of the most challenging parts in medical image segmentation due to its complex anatomical structure and complex surrounding environment.Aiming at the above problems,a deep learning segmentation model combining dual decoding and global attention upsampling module(DGANet)is proposed.The model consists of an encoder and two decoders,where the latter realizes the full utilization of different depth feature information.The model applies the global attention upsampling module and high-level rich semantic information to guide the low-level selection of more accurate feature information.The data set provided by Changhai Hospital was used for experiments.The results showed that the average Dice similarity coefficient was 86.28%,Intersection-over-Union(IoU)was 0.77,and Hausdorff distance(HD)was 7.7 mm.The data confirmed the clinical value of this model in segmenting pancreatic cystic tumors.

关 键 词：医学图像分割 DGANet 深度学习 胰腺 

分 类 号：O482.53[理学—固体物理]