激光与光电子学进展, 2018, 55 (8): 081001, 网络出版: 2018-08-13
基于迁移学习和深度卷积神经网络的乳腺肿瘤诊断系统 下载: 1502次
Breast Cancer Diagnosis System Based on Transfer Learning and Deep Convolutional Neural Networks
图像处理 乳腺肿瘤诊断 图像分类 深度卷积神经网络 计算机辅助诊断系统 迁移学习 image processing breast tumor diagnosis image classification deep convolutional neural networks computer-aided diagnostic system transfer learning
摘要
乳腺肿瘤计算机辅助诊断(CAD)系统在医学检测和诊断中的应用日益重要。为了区分核磁共振图像(MRI)中肿瘤与非肿瘤,利用深度学习和迁移学习方法,设计了一种新型乳腺肿瘤CAD系统:1)对数据集进行不平衡处理和数据增强;2)在MRI数据集上,利用卷积神经网络(CNN)提取CNN特征,并利用相同的支持向量机分类器,计算每层CNN的特征图的分类F1分数,选取分类性能最高的一层作为微调节点,其后维度较低层为连接新网络节点;3)在选取的网络接入节点,连接新设计的两层全连接层组成新的网络,利用迁移学习,对新网络载入权重;4)采用固定微调节点前的网络层不可训练,其余层可训练的方式微调。分别基于深度卷积网络(VGG16)、Inception V3、深度残差网络(ResNet50)构建的CAD系统,性能均高于主流的CAD系统,其中基于VGG16和ResNet50搭建的系统性能突出,且二次迁移可以提高VGG16系统性能。
Abstract
Breast cancer computer-aided diagnosis (CAD) system is playing more and more important role in medical detection and diagnosis. In order to classify tumor and non-tumor in magnetic resonance imaging (MRI), a novel breast cancer CAD system based on deep learning and transfer learning is designed. First, we balance the imbalanced data sets and use data augmentation to deal with it. Then, we use the convolutional neural network (CNN) to extract CNN features from MRI data sets, use the same support vector machine to evaluate the feature extraction abilities of different layers, and select the highest F1 score layer as the node of fine-tuning, the layers behind it, which has relatively low dimension as the node of connection of new networks. Next, we select the newly designed fully-connected layers with two layers to form a new network, and use transfer learning to load weights on the new network. At last, we freeze the layers before the node of fine-tuning, while other layers can be trained in the fine-tuning procedure. The CAD systems are built on three CNN networks, including VGG16, Inception V3, and ResNet50. The effects of the system based on VGG16 and ResNet50 have the best performance, and twice transfer learning can improve the performance of VGG16 network system.
褚晶辉, 吴泽蕤, 吕卫, 李喆. 基于迁移学习和深度卷积神经网络的乳腺肿瘤诊断系统[J]. 激光与光电子学进展, 2018, 55(8): 081001. Chu Jinghui, Wu Zerui, Lü Wei, Li Zhe. Breast Cancer Diagnosis System Based on Transfer Learning and Deep Convolutional Neural Networks[J]. Laser & Optoelectronics Progress, 2018, 55(8): 081001.