查询结果:   郑群花,段慧芳,沈尧,刘娟,袁静萍.基于卷积神经网络和迁移学习的乳腺癌病理图像分类[J].计算机应用与软件,2018,35(7):237 - 242.
中文标题
基于卷积神经网络和迁移学习的乳腺癌病理图像分类
发表栏目
图像处理与应用
摘要点击数
922
英文标题
BREAST CANCER HISTOLOGICAL IMAGE CLASSIFICATION BASED ON CONVOLUTIONAL NEURAL NETWORK AND TRANSFER LEARNING
作 者
郑群花 段慧芳 沈尧 刘娟 袁静萍 Zheng Qunhua Duan Huifang Shen Yao Liu Juan Yuan Jingping
作者单位
武汉大学计算机学院 湖北 武汉 430072 武汉大学人民医院病理科 湖北 武汉 430060    
英文单位
School of Computer Science, Wuhan University,Wuhan 430072, Hubei, China Department of Pathology, Renmin Hospital of Wuhan University, Wuhan 430060, Hubei, China    
关键词
乳腺癌病理图像 卷积神经网络 图像分块 多数投票算法 迁移学习
Keywords
Breast cancer histological image CNN Image patch Majority voting algorithm Transfer learning
基金项目
国家自然科学基金项目(61272274)
作者资料
郑群花,硕士生,主研领域:医学图像处理,深度学习。段慧芳,硕士生。沈尧,硕士生。刘娟,教授。袁静萍,博士。 。
文章摘要
乳腺癌已经成为导致女性死亡最常见的和发病率最高的恶性肿瘤。对HE染色的乳腺癌病理图像的自动分类具有重要的临床意义。针对目前存在的深度卷积神经网络只将图像分为良性和恶性两类,同时对于高分辨率图像处理具有局限性的不足,采用了以AlexNet为架构的卷积神经网络模型将图像分为乳腺导管原位癌、乳腺浸润性导管癌、乳腺纤维腺瘤和乳腺增生四类。对于高分辨率图像,采用图像分块的思想,将每块的分类结果利用多数投票算法进行整合,整合结果作为该图像的分类结果。同时,为了避免因乳腺癌病理图像标记样本过少带来的过拟合问题,采用了迁移学习和数据增强的方法。实验结果表明,该模型识别率达到了99.74%。
Abstract
Breast cancer has been the most common and morbid malignant tumor that causes female death. The auto-classification of HE-stained breast cancer histological image has a significant clinical value. In view of the existing convolutional neural networks(CNN), images are only divided into two classes, benign and malignant. At the same time, they have limitations for high-resolution image processing. This paper proposed a CNN model based on Alexnet to classify the images into four classes, ductal carcinoma in situ, invasive ductal carcinoma, breast fibroadenoma and breast hyperplasia. For high-resolution images, this paper divided them into patches. The classification results of each patch were integrated by majority voting algorithm, and the integration result was invoked as the classification result of the image. Besides, transfer learning and data augmentation were adopted in order to avoid the overfitting caused by the limitation of labeled training samples. The experimental results demonstrate that the model recognition rate reaches 99.74%.
下载PDF全文