Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (7): 1243-1250.DOI: 10.3778/j.issn.1673-9418.1907053

Previous Articles     Next Articles

Method for Intestinal Polyp Segmentation by Improving DeepLabv3+ Network

WANG Yagang, XI Yiyuan, PAN Xiaoying   

  1. 1. School of Computer Science, Xi'an University of Posts & Telecommunications, Xi'an 710121, China
    2. Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing, Xi'an University of Posts & Tele-communications, Xi'an 710121, China
  • Online:2020-07-01 Published:2020-08-12

改进DeepLabv3+网络的肠道息肉分割方法

王亚刚郗怡媛潘晓英   

  1. 1. 西安邮电大学 计算机学院,西安 710121
    2. 西安邮电大学 陕西省网络数据分析与智能处理重点实验室,西安 710121

Abstract:

In order to enhance the detection rate of polyp of intestine under colonoscopy, an improved DeepLabv3+ network method for intestinal polyp segmentation is proposed. In the data preprocessing stage, using the nonlinear filtering characteristics of the median filter to remove the image reflection area, and Grab Cut algorithm is combined to pre-extract the polyp area. Coarse segmentation results of polyp location are obtained, which are superimposed with the original drawing to reinforce the signal strength of polyp location. In terms of network structure, this paper introduces the optimal dense prediction cell obtained through neural architecture search into DeepLabv3+ network, uses 3-layer depth separable convolution to gradually acquire segmentation results in the decoder part, so as to reduce incomplete segmentation in the segmentation process. In the experiment, through training and testing of CVC-ClinicDB data set, the average joining and merging ratio, Dice coefficient, sensitivity, precision and F1 value are used as judgment standard. The mean intersection over union reaches 0.947, and the other 4 indexes are all higher than 0.935. The experimental results show that compared with the existing methods, the proposed method in this paper improves the accuracy of intestinal polyp image segmentation to a certain extent, which can be used for reference in the processing and analysis of intestinal polyp images by deep learning.

Key words: improved DeepLabv3+, intestinal polyp, neural architecture search, incomplete segmentation

摘要:

为了提高结肠镜下肠道息肉检测率,提出了一种改进DeepLabv3+网络的肠道息肉分割方法。在数据预处理阶段,利用中值滤波的非线性滤波特性去除掉图像反光区域,并结合Grab Cut算法对息肉区域进行预提取,得到息肉位置的粗分割结果,将其与原图叠加以增强息肉位置的信号强度。在网络结构上,将通过神经架构搜索得到的最优密集预测单元引入DeepLabv3+网络,并在解码器部分采用3层深度可分离卷积逐步获取分割结果,减少分割过程中不完全分割的情况。实验通过对CVC-ClinicDB数据集进行训练和测试,以平均交并比、Dice系数、敏感度、精确率以及F1值作为评判标准,其中平均交并比达到0.947,其余4项指标也均高于0.935。实验结果表明提出的方法与现有方法相比,对肠道息肉图像分割在精度上有一定提升,对深度学习在肠道息肉图像的处理和分析具有借鉴意义。

关键词: 改进DeepLabv3+, 肠道息肉, 神经架构搜索, 不完全分割