作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (8): 180-186. doi: 10.19678/j.issn.1000-3428.0062023

• 图形图像处理 • 上一篇    下一篇

基于内卷U-Net的医学图像分割模型

林志洁1, 郑秋岚2, 梁涌3, 邢卫4   

  1. 1. 浙江科技学院 信息与电子工程学院, 杭州 310018;
    2. 杭州医学院 食品科学与工程学院, 杭州 310013;
    3. 丽水市第二人民医院 精神病科, 浙江 丽水 323000;
    4. 浙江大学 计算机科学与技术学院, 杭州 310027
  • 收稿日期:2021-07-08 修回日期:2021-10-11 发布日期:2021-10-18
  • 作者简介:林志洁(1980-),男,副教授、博士,主研方向为图像处理;郑秋岚,硕士;梁涌,高级政工师;邢卫,副教授、博士。
  • 基金资助:
    浙江省公益基础项目(LGF19HO90025);浙江省医药卫生项目(2018KY037)。

Medical Image Segmentation Model Based on Involution U-Net

LIN Zhijie1, ZHENG Qiulan2, LIANG Yong3, XING Wei4   

  1. 1. School of Information and Electronic Engineering, College of Science and Technology of Zhejiang, Hangzhou 310018, China;
    2. School of Food Science and Engineering, Hangzhou Medical College, Hangzhou 310013, China;
    3. Psychiatry Department, Lishui Second People's Hospital, Lishui, Zhejiang 323000, China;
    4. College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China
  • Received:2021-07-08 Revised:2021-10-11 Published:2021-10-18

摘要: 图像分割技术的主要对象为自然图像和医学图像,相对于自然图像而言,医学图像的语义分割通常需要较高的精度以进行下一步的临床分析、诊断和规划治疗。目前用于医学图像语义分割的深度神经网络模型由于仅考虑位置的平移不变性,存在局部感受野较小、无法表达长范围依赖关系的问题。设计一种面向医学图像的分割模型,基于内卷U-Net网络,使用内卷操作代替传统的卷积操作,并将内卷结构作为基本的网络结构,提升模型对医学图像局部特征的学习能力。在模型的瓶颈层引入注意力机制模块来学习图像长范围的依赖关系,以提高医学图像语义分割的精度。在肺部CT数据集上的实验结果表明,该模型的Dice系数为0.998,较基于卷积神经网络的分割模型约提高5%,并且大幅缩短Hausdorff距离,具有更高的分割准确度以及较好的稳健性。

关键词: 图像分割, 内卷网络, 图像处理, 注意力机制, U-Net网络

Abstract: The main objects of image segmentation technology are natural images and medical images.Compared with that of natural images, semantic segmentation of medical images usually requires high accuracy for the subsequent steps of clinical analysis, diagnosis, and treatment planning.At present, the depth neural network model used to semantically segment medical images only considers the translation invariance of position, which features an insufficiently large local receptive field, leaving no way to express long-range dependence.Therefore, medical image segmentation model is proposed in this paper.Based on the Involution U-Net network, the involution operation replaces the traditional convolution operation, and the involution structure is adopted as the basic network structure to tailor the model's learning ability to the local features of medical images.To improve the model's learning ability, Besides, the attention mechanism module is introduced into the bottleneck layer to learn the long-range dependency of images.Experimental results on the lung CT dataset show that the model's Dice coefficient is 0.998, which is approximately 5% higher than that of the current segmentation model based on a convolutional neural network.In addition, this model greatly cuts the Hausdorff distance, apart from achieving higher segmentation accuracy and better robustness.

Key words: image segmentation, involution network, image processing, attention mechanism, U-Net network

中图分类号: