基于小波分解卷积神经网络的病理图像分类  

Pathological Image Classification Based on Wavelet Decomposition CNN

作　　者：丁偕[1] 崔浩阳 张敬谊[1] DING Xie;CUI Hao-Yang;ZHANG Jing-Yi(Data Center,Wonders Information Co.Ltd.,Shanghai 201112,China)

机构地区：[1]万达信息股份有限公司数据中心,上海201112

出　　处：《计算机系统应用》2021年第9期322-329,共8页Computer Systems & Applications

基　　金：上海市科技创新项目(18511102700);上海市人工智能创新发展专项(2019-RGZN-01017)。

摘　　要：组织病理图像分析是癌症诊断的"金标准",在患者的预后治疗中起到至关重要的作用.目前在AI医学影像领域,利用CNN(Convolutional Neural Network)网络对数字病理图像的分类已经成为研究热点.但是传统CNN网络中广泛使用最大/平均池化(Max/Average pooling)模块,不可避免的丢失了大量病理图像中的特征信息,造成分类准确率低且模型不易收敛.因此,本文提出一种基于小波分解卷积神经网络的病理图像分类方法(Wavelet Decomposition Convolutional Neural Networks,WDCNN),该方法能够使传统CNN模型学习到频域信息,它将小波变换的多尺度分析引入到CNN模型中,利用小波分解代替传统的池化层,相比于最大值和平均值池化减少了特征的丢失.鉴于空域与频域具有不同的特性,将小波分解后的高频分量通过捷径连接的方式添加到下一层,弥补了在池化过程中丢失的细节特征信息.本文在Camelyon16数据集上评估了不同的池化方法和不同小波基函数在病理图像分类方面的性能.根据实验结果表明,引入小波分解的CNN模型能够提升网络的分类准确率.Histopathological image analysis is the"gold standard"for cancer diagnosis,which plays an important role in the prognosis and treatment of patients.Currently,in the field of AI medical imaging,the classification of pathological images based on Convolutional Neural Network(CNN)has become a research hotspot.However,the Max/Average pooling module is widely used in traditional CNNs,which inevitably lose massive feature information in pathological images,resulting in low classification accuracy and difficult model convergence.Therefore,this study proposes a pathological image classification method based on Wavelet Decomposition Convolutional Neural Network(WDCNN).This method can make the traditional CNN model learn the frequency domain information.It introduces the multi-scale analysis of wavelet transform into a CNN model and uses wavelet decomposition to replace the traditional pooling layer,which reduces the loss of features compared with max and average pooling.In view of different characteristics of the space domain and the frequency domain,the high-frequency components after wavelet decomposition are added to the next layer through shortcut connections to make up for the detailed feature information lost in the pooling process.This paper evaluates the performance of different pooling methods and different wavelet basis functions in pathological image classification on the Camelyon16 dataset.According to the experimental results,the CNN model integrated with wavelet decomposition can improve the classification accuracy of the network.

关 键 词：病理图像 卷积神经网络(CNN) 小波分解 捷径连接 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] R-05[医药卫生]