南方医科大学学报

南方医科大学学报 ›› 2021, Vol. 41 ›› Issue (11): 1616-1622.doi: 10.12122/j.issn.1673-4254.2021.11.04

• • 上一篇    下一篇

基于激活层前置压缩激励残差网络的早期胃癌筛查算法

张晓玥，王永雄，张佳鹏，孙洪鑫，王 东，陈 羽，周 志   

  1. 上海理工大学光电信息与计算机工程学院，上海 200093；上海长海医院消化内科，上海 200433；南方医科大学第七附属医院消化内科，广东 佛山 528244
  • 出版日期:2021-11-20 发布日期:2021-12-10

Screening of early gastric cancer using Pre-Activation Squeeze-and-Excitation ResNet

ZHANG Xiaoyue, WANG Yongxiong, ZHANG Jiapeng, SUN Hongxin, WANG Dong, CHEN Yu, ZHOU Zhi   

  1. School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China; Department of Gastroenterology, Changhai Hospital, Shanghai 200433, China; Seventh Affiliated Hospital of Southern Medical University, Foshan 528244, China
  • Online:2021-11-20 Published:2021-12-10

RichHTML

19

PDF

470

摘要： 目的 研究基于激活层前置压缩激励残差网络（PASE-ResNet）的快速和准确的早期胃癌筛查算法。方法 构建一个基于激活层前置压缩激励残差网络的早期胃癌筛查算法。为聚焦任务相关的图像区域，提升模型的特征表达能力，将压缩激励模块（SE）与激活层前置残差网络（PreAct-ResNet）中的残差模块相结合，提高与任务相关的特征通道权重。此外，为提高早期胃癌的筛查性能，提出“局部筛查+全局滑窗”的策略，经数据扩充后得到数据集子图18 400幅。利用PASE-ResNet模型通过滑动窗口的方式对胃镜图像进行检测，获得了精细的筛查结果。结果 本文提出的模型在早期胃癌筛查中取得了98.03%的准确率、98.96%的灵敏度和96.52%的特异性值。结论 本文提出的基于激活层前置压缩激励残差网络，达到了较好的筛查精度，有望在临床中辅助医生快速诊断。

关键词: 早期胃癌筛查；残差网络；注意力机制

Abstract: Objective To propose a quick and accurate method for screening early gastric cancer based on Pre-Activation Squeeze- and-Exception ResNet (PASE-ResNet) gastroscopy images in limited labeled data sets. Methods We developed an algorithm based on Pre-Activation Squeeze- and-Exception ResNet for early gastric cancer screening. To focus on the task-related image region and enhance the feature expression ability of model, we combined the Squeeze-and-Exception (SE) module with the residual module in PreAct-ResNet to adjust the weight of the feature channel. The strategy of local screening + global sliding window was adopted to improve the performance of early cancer screening. After data expansion, 18 400 set subgraphs were obtained, and the gastroscopy images were examined using the PASE-ResNet model by sliding window. Results The results of experiments showed that the proposed algorithm had good performance for screening early gastric cancer with an accuracy of 98.03% , a sensitivity of 98.96% and a specificity of 96.52% . Conclusion The PASE-ResNet can achieve a high accuracy for screening early gastric cancer.

Key words: screening of early gastric cancer; residual network; attention mechanism