Breast Cancer Image Classification via Multi-Network Features and Dual-Network Orthogonal Low-Rank Learning

Abstract

Histopathological image analysis is an important technique for early diagnosis and detection of breast cancer in clinical practice. However, it has limited efficiency and thus the detection of breast cancer is still an open issue in medical image analysis. To improve the early diagnostic accuracy of breast cancer and reduce the workload of doctors, we devise a classification framework based on histology images by combining deep learning with machine learning methodologies in this paper. Specifically, we devise a multi-network feature extraction model by using pre-trained deep convolution neural networks (DCNNs), develop an effective feature dimension reduction method and train an ensemble support vector machine (E-SVM). First, we preprocess the histological images via scale transformation and color enhancement methods. Second, the multi-network features are extracted by using four pre-trained DCNNs (e.g., DenseNet-121, ResNet-50, multi-level InceptionV3, and multi-level VGG-16). Third, a feature selection method via dual-network orthogonal low-rank learning (DOLL) is further developed for performance boosting and overfitting alleviation. Finally, an E-SVM is trained via fused features and voting strategy to perform the classification task, which classifies the images into four classes (i.e., benign, in situ carcinomas, invasive carcinomas, and normal). We evaluate the proposed method on the public ICIAR 2018 Challenge dataset of histology images of breast cancer and achieve a high classification accuracy of 97.70%. Experimental results show that our method can achieve quite promising performance and outperform state-of-the-art methods.