基于融合多网络深层卷积特征和稀疏双关系正则化方法的乳腺癌图像分类研究

Breast Cancer Image Classification Based on Fusion Multi-Network Deep Convolution Features and Sparse Double Relation Regularization Method

乳腺癌是全球女性癌症死亡的主要原因之一。现有诊断方法主要是医生通过乳腺癌观察组织病理学图像进行判断,不仅费时费力,而且依赖医生的专业知识和经验,使得诊断效率无法令人满意。针对以上问题,设计基于组织学图像的深度学习框架,以提高乳腺癌诊断准确性,同时减少医生的工作量。开发一个基于多网络特征融合和稀疏双关系正则化学习的分类模型:首先,通过子图像裁剪和颜色增强进行乳腺癌图像预处理;其次,使用深度学习模型中典型的3种深度卷积神经网络(Inception V3、Res Net-50和VGG-16),提取乳腺癌病理图像的多网络深层卷积特征并进行特征融合;最后,通过利用两种关系("样本-样本"和"特征-特征"关系)和lF正则化,提出一种有监督的双关系正则化学习方法进行特征降维,并使用支持向量机将乳腺癌病理图像区分为4类—正常、良性、原位癌和浸润性癌。实验中,通过使用ICIAR 2018公共数据集中的400张乳腺癌病理图像进行验证,获得93%的分类准确性。融合多网络深层卷积特征可以有效地捕捉丰富的图像信息,而稀疏双关系正则化学习可以有效降低特征冗余并减少噪声干扰,有效地提高模型的分类性能。

Breast cancer is one of the leading causes of cancer death worldwide.Existing diagnostic methods are mainly dependent on the observation with histopathological images,which is laborious,time-consuming,and relies on the doctor’s professional knowledge and experience,making the diagnosis efficiency unsatisfactory.In view of these problems,this paper aimed to improve the breast cancer diagnostic accuracy and reduce the workload of doctors by devising a deep learning framework based on histological image.Specifically,this paper developed a classification model based on multi-network feature fusion and sparse double-relation regularized learning.First,the breast cancer pathological images were preprocessed by sub-image clipping and color enhancement.Then,three deep convolutional neural networks(InceptionV3,Res Net-50,and VGG-16)typical of deep learning model were used to extract multi-network deep convolution features of breast cancer pathological images.Third,by using two relations("sample-sample"relation and"feature-feature"relation)and lF regularization,we proposed a supervised double relation regularization learning method to reduce feature dimension.Support vector machines was used to distinguish breast cancer pathological images into four categories:normal,benign,carcinoma in situ,and invasive carcinoma.In the experiment,by using 400 breast cancer pathological images in the ICIAR 2018 public data set to verify the proposed method,93%classification accuracy was obtained.Results showed that multi-network deep convolution fusion features could effectively capture rich image information,and sparse dual-relation regularization learning could effectively reduce feature honor and reduce noise interference,which will effectively improve the classification performance of the model.