Deep Co-Training Active Learning for Mammographic Images Classification

Abstract

Breast cancer becomes one of the most serious women's diseases. With help of early screen, the deep learning based computer-aided diagnosis (CAD) could reduce the mortality rate, which has become increasingly concerned. Building a deep learning based CAD system requires a large labeled dataset, however the labeling procedure is highly time consuming and labor-intensive. The active learning algorithm is applied to reduce the labeling cost by selecting the most valuable sample, which improve the model performance with smaller amount of data and limited annotation. However, the informative unlabeled data will be discarded. To make use of unlabeled data, the co-training approach is a suitable option apply in breast mammogram classification. We propose the co- training active learning (COAL) framework combine with the co-training and active learning which not only leverage the unlabeled data and import the annotation efficiency but also simultaneously utilizes two-view images to improve the classification performance.The experimental results proved the effectiveness and efficiency of the framework, where only 30% labeled datasets could achieve a comparable performance to the fully-supervised learning framework. Besides, the proposed algorithm fully utilizes the breast images two-view information, which could improve 2% prediction accuracy.