基于卷积神经网络的乳腺癌病理图像分类方法  

BREAST CANCER IMAGE CLASSIFICATION METHOD BASED ON CONVOLUTIONAL NEURAL NETWORK

作　　者：龚安[1] 吕秀明 Gong An;LüXiuming(College of Computer Science and Technology,China University of Petroleum,Qingdao 266580,Shandong,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580

出　　处：《计算机应用与软件》2023年第6期133-139,198,共8页Computer Applications and Software

基　　金：国家科技重大专项(2017ZX05013-001);中石油重大科技项目(ZD2019-183-004);中央高校基本科研业务费专项资金项目(20CX05019A)。

摘　　要：针对单一卷积神经网络在乳腺癌识别分类准确率不高、研究集中二元分类等问题,基于深度学习提出一种多模型融合机制方法。对乳腺癌组织学图像进行预处理,通过数据增强方法缓解数据集较少且分布不均匀问题;使用六个CNN通过迁移学习策略进行训练,提取多网络特征并保存,通过验证集损失率选出最优两个模型ResNet50和Inception_v3进行融合;实现病理图像在不同放大倍数下的多级分类。实验结果表明,模型融合后在患者级别和图像级别分类准确率最高达到94.18%、94.12%,优于单一网络、传统机器学习方法和现有基于深度学习二元分类方法,说明该网络有助于乳腺癌病理图像的分类研究。Aimed at the problems of low accuracy of single convolutional neural network in breast cancer recognition and classification and binary classification in research focus,a multi-model fusion mechanism method based on deep learning is proposed.The breast cancer histological image was preprocessed,and the problem of less data sets and uneven distribution was alleviated through data enhancement.Six CNN models were used to train through the transfer learning strategy,and the multi-network features were extracted and saved.ResNet50 and Inception_v3 with the smallest loss function on the verification set were selected for fusion.It realized multi-level classification of pathological images under different magnification.The experimental results show that the model s accuracy at the level of the image and patient after fusion is up to 94.18%and 94.12%,which are higher than single network,traditional machine learning methods and binary classification methods based on deep learning.The network is helpful for the classification of breast cancer pathological images.

关 键 词：乳腺癌识别 模型融合 卷积神经网络 迁移学习 数据增强 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]