基于联合训练的分类器的乳腺癌图像分类  

Breast Cancer Image Classification Based on Joint Training Classifier

作　　者：张晋凯 高翔[1] 王鹏[1] 白艳萍[1] 梅银珍[1] ZHANG Jinkai;GAO Xiang;WANG Peng;BAI Yanping;MEI Yinzhen(School of Mathematics,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学数学学院,太原030051

出　　处：《计算机测量与控制》2023年第5期228-234,共7页Computer Measurement &Control

基　　金：国家自然科学基金项目(61774137);山西省基础研究计划资助项目(202103021224195,202103021224212,202103021223189,20210302123019);山西省回国留学人员科研项目(2020-104,2021-108)。

摘　　要：利用机器学习的乳腺癌组织病理图像诊断节省了大量的人力物力,因此提高乳腺癌组织病理图像识别准确率有很好的现实意义;针对单一分类器和集成学习分类器模型观测域有限容易陷入局部最优的问题,提出一种基于联合训练的分类器模型;通过单一分类器相互影响扩大观测感知域来寻找损失最小的估计点,根据估计点来迭代优化超参数进而联合训练出拟合性能最好的分类器,这样既汲取不同分类器模型的可取之处来增强泛化能力,又加大了模型观测域在可以更快的得到全局最优的同时提升了识别准确率;实验表明,提出的联合训练的分类器能够提升乳腺癌组织病理学图像的分类性能,在不同放大倍数40×、100×、200×、400×下图像良恶性分类准确率分别为99.67%、98.08%、99.01%、96.34%。The histopathological image diagnosis of breast cancer using machine learning saves a lot of manpower and material resources,so it is of good practical significance to improve the histopathological image recognition accuracy of breast cancer.Aiming at the problem that the observation domain of single classifier and ensemble learning classifier model is limited and easy to fall into local optimality,a classifier model based on joint training is proposed.Through the interaction of single classifier,the observation perception domain is expanded to find the estimated point with the least loss,and the hyperparameters are iteratively optimized according to the estimated point and then jointly trained the classifier with the best fitting performance,which not only absorbs the desirability of different classifier models to enhance the generalization ability,but also increases the model observation domain to obtain the global optimal faster while improving the recognition accuracy.The experimental results show that the proposed jointly training classifier can improve the classification performance of the breast cancer histopathological images,and the accuracy of benign and malignant classification at different magnifications of 40×,100×,200×and 400×is 99.67%,98.08%,99.01%and 96.34%,respectively.

关 键 词：乳腺癌图像 分类器模型 联合训练 深度学习 机器学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]