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 ›› 2020, Vol. 46 ›› Issue (9): 248-253. doi: 10.19678/j.issn.1000-3428.0056011

• 图形图像处理 • 上一篇    下一篇

改进U-Net在喉白斑病灶分割中的应用

吉彬1, 任建君2, 郑秀娟1, 谭聪3, 吉蓉4, 赵宇2, 刘凯1   

  1. 1. 四川大学 电气工程学院 自动化系, 成都 610065;
    2. 四川大学华西医院 耳鼻咽喉-头颈外科, 成都 610041;
    3. 成都医学院第一附属医院 耳鼻咽喉头颈外科, 成都 610500;
    4. 西安医学院 临床医学院, 西安 710021
  • 收稿日期:2019-09-16 修回日期:2019-10-23 发布日期:2019-11-05
  • 作者简介:吉彬(1993-),男,硕士研究生,主研方向为医学图像处理;任建君,讲师、博士;郑秀娟,副教授、博士;谭聪,学士;吉蓉,硕士研究生;赵宇、刘凯,教授、博士。
  • 基金资助:
    国家自然科学基金(81201146)。

Application of Improved U-Net in Segmentation of Laryngeal Leukoplakia Lesion

JI Bin1, REN Jianjun2, ZHENG Xiujuan1, TAN Cong3, JI Rong4, ZHAO Yu2, LIU Kai1   

  1. 1. Department of Automation, College of Electrical Engineering, Sichuan University, Chengdu 610065, China;
    2. Department of Otolaryngology-Head and Neck Surgery, West China Hospital of Sichuan University, Chengdu 610041, China;
    3. Department of Otolaryngology Head and Neck Surgery, The First Affiliated Hospital of Chengdu Medical College, Chengdu 610500, China;
    4. College of Clinical Medicine, Xi'an Medical University, Xi'an 710021, China
  • Received:2019-09-16 Revised:2019-10-23 Published:2019-11-05

摘要: 喉白斑属于癌前组织病变,准确检测该病灶对癌变预防和病变治疗至关重要,但喉镜图像中病灶边界模糊且表面反光导致其不易分割。为此,提出一种基于U-Net的多尺度循环卷积神经网络(MRU-Net)进行喉白斑病灶分割。通过对比度受限的自适应直方图均衡化技术增强喉镜图像,利用平均池化构建图像金字塔并将其作为U型网络多尺度输入,同时使用多尺度卷积和递归卷积层代替编码与解码单元卷积层改进网络结构,采用多尺度输出层生成不同尺度特征图并对各层求均值得到最终输出结果。实验结果表明,MRU-Net的F1值、Jaccard相似度和平均交并比分别为0.784 3、0.661 1和0.826 9,与U-Net、M-Net等传统网络相比,该网络对喉白斑病灶分割更准确,能够得到精度更高的病灶轮廓。

关键词: 卷积神经网络, 病灶分割, 喉镜图像, 循环卷积, 喉白斑

Abstract: Laryngealleukoplakia is a kind of precancerous tissue lesion.Accurate detection of the lesion is very important for the prevention and treatment of cancer.However,the edge of the lesion in laryngoscope images which is blurred and the surface reflection make it difficult to segment.Therefore,this paper proposes a U-Net-based multi-scale recurrent Convolutional Neural Network(CNN)(MRU-Net) to segment laryngeal leukoplakia lesion.The network uses Contrast-Limited Adaptive Histogram Equalization(CLAHE) technology to enhance laryngoscope images,and the image pyramid is constructed by average pooling as the multi-scale input of U-shaped network.At the same time,the multi-scale convolution and Recursive Convolution Layer (RCL) are used to replace the convolution layer of coding and decoding units to improve the network structure.The multi-scale output layer is used to generate different scales of feature maps,and each layer is averaged to generate the final result.Experimental results show that the F1 value,Jaccard Similarity(JS) and Mean Intersection over Union(MIoU) of MRU-Net were 0.784 3,0.661 1 and 0.826 9,respectively.Compared with traditional medical image segmentation methods such as U-Net and M-Net,the proposed network is more accurate in segmentation of laryngeal leukoplakia lesion,and the precision of obtained lesion contour is higher.

Key words: Convolutional Neural Network(CNN), lesion segmentation, laryngoscopy image, recurrent convolution, laryngeal leukoplakia

中图分类号: