Meta-learning Based Breast Abnormality Classification on Screening Mammograms

Abstract

General breast cancer detection contains two steps, the breast abnormality classification, and the diagnostic classification. The determination of the abnormality contributes further to the following steps, and computational technologies can aid in the process. A lot of machine learning methods have been applied to automate the detection. However, most of them focus on the diagnostic classification and the breast abnormality classification only attracts little attention. The insufficient size of public mammogram datasets also limits the performance of many machine learning algorithms. Considering the importance of breast abnormality classification and the shortage of public large-scale medical datasets, we proposed a meta-learning-based breast abnormality classification method. Our model referred to the latest work of meta-learning-based image classifier and modified it. Specifically, we applied the idea of meta-learning to retrain a pretrained embedding neural network in order to adapt its feature extraction ability to the CBIS-DDSM dataset [1]. The dataset contains two types of abnormal breast mammograms, mass and calcification, and each type is made of two categories of medical images, full mammograms, and ROI [2]. The application of the data augmentation techniques and the idea of meta-learning helped to deal with the insufficient training sample problem and showed a final accuracy of 76%, which beat the 71% accuracy reached by a neural network baseline model.