采用密集多尺度卷积网络的胰腺分割方法  

Pancreas Segmentation Method Based on Dense Multi-scale Convolutional Network

作　　者：张学鹏 王远军[1] ZHANG Xue-peng;WANG Yuan-jun(Institute of Medical Imaging Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学医学影像工程研究所,上海200093

出　　处：《小型微型计算机系统》2022年第12期2614-2619,共6页Journal of Chinese Computer Systems

基　　金：上海市自然科学基金面上项目(18ZR1426900)资助。

摘　　要：胰腺是人体内重要的消化器官,受个体年龄、健康状况等因素的影响,它的形状、尺寸和位置可能会发生较大变化.胰腺自动分割一直以来是医学图像分析和计算机辅助诊断领域一个具有挑战性的问题.近年来,深度学习在医学图像分割领域上得到了广泛的应用,本文提出了一种密集多尺度卷积网络(Dense multi-scale convolutional networks,DMC-net)以用于进行胰腺的自动分割.本文将多层图像作为网络输入,采用密集卷积和密集多尺度卷积连接代替了U-net的常规卷积和长跳跃连接,此外在训练过程中本文还采用了边界损失函数对胰腺的形状进行约束.在NIH胰腺公开数据集上的结果表明,文中方法的分割结果Dice系数可以达到86.19%,证明了本文提出的胰腺分割方法的有效性.The pancreas is an important digestive organ in the human body,and its shape,size,and position may undergo major changes due to factors such as individual age and health status.Automatic pancreas segmentation has always been a challenging problem in the field of medical image analysis and computer-aided diagnosis.In recent years,deep learning has been widely used in the field of medical image segmentation.We proposed Dense multi-scale convolutional networks(DMC-net)for automatic pancreas segmentation.We take multi-layer images as the network input,and dense convolution block and dense multi-scale convolution connections were used to replace the the traditional convolution block and long skip connection of U-net.In addition,we used a kind of boundary loss function to constraint the shape of the pancreas during the training process.The results on the NIH pancreas public datasets show that the Dice coefficient of the segmentation result of our method in this paper can reach 86.19%,which proves the effectiveness of the pancreas segmentation method proposed in this paper.

关 键 词：胰腺分割 卷积网络 密集卷积 多尺度卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]