基于深度迁移学习的肠道息肉癌变风险评估  

Risk assessment of intestinal polyp canceration based on deep transfer learning

作　　者：程珊 叶然 曹婧 姚佳锋 李胜[1] 何熊熊[1] CHENG Shan;YE Ran;CAO Jing;YAO Jiafeng;LI Sheng;HE Xiongxiong(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023;Lishui People’s Hospital,Lishui 323400)

机构地区：[1]浙江工业大学信息工程学院,杭州310023 [2]丽水市人民医院,丽水323400

出　　处：《高技术通讯》2023年第6期620-629,共10页Chinese High Technology Letters

基　　金：国家自然科学基金(61873239);浙江省教育厅一般科研(Y202043133)资助项目。

摘　　要：针对肠道息肉分类中不同类型息肉难以区分的问题,提出了一种基于深度迁移学习的肠道息肉癌变风险评估方法。该方法首先在主干网络中引入通道注意力进行息肉特征提取;然后将任务网络注意力图分支输出的注意力图加权到主干网络输出的特征图上,获取更精确的病灶特征;最后将加权后的特征图输入到任务网络的感知分支,使用2个分支的损失函数以端到端的方式优化模型,以进行息肉癌变风险评估。本文方法对数据集进行了增强,并结合迁移学习进一步提升分类效果,最优模型的灵敏度、特异度、精准度、F1分数及曲线下方的面积(AUC)分别为99.38%、98.15%、98.17%、98.77%和0.9996。实验结果表明,本文方法能有效地对肠道息肉癌变风险进行评估,性能优于该领域流行的深度学习方法。To accurately distinguish different types of polyps,a novel risk assessment method based on deep transfer learning is proposed.Firstly,a channel attention is introduced into the backbone network to extract the polyp fea-tures.Then,the attention map from the attention map branch of task network is weighted to the feature map from the backbone to obtain more accurate lesion features.Finally,the weighted feature map is fed into the perception branch of the task network,and loss functions of the two branch is used to optimize the model in an end-to-end manner to determine the canceration risk of polyps.The method enhances the established dataset,and employ transfer learning to further improve the classification results.The sensitivity,specificity,precision,F1-Score and area under the ROC curve(AUC)of the optimal model are 99.38%,98.15%,98.17%,98.77%,0.9996,re-spectively.The experimental results show that this method can effectively evaluate the canceration risk of intestinal polyps with better performance compared with the popular deep learning methods.

关 键 词：深度学习 迁移学习 通道注意力 注意力图 图像分类 肠道息肉 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] R735.3[医药卫生—肿瘤]