Breast mass detection from the digitized X-ray mammograms based on the combination of deep active learning and self-paced learning

Abstract

Abstract Breast mass detection is a challenging task in mammogram, since mass is usually embedded and surrounded by various normal tissues with similar density. Recently, deep learning has achieved impressive performance on this task. However, most deep learning methods require large amounts of well-annotated datasets. Generally, the training datasets is generated through manual annotation by experienced radiologists. However, manual annotation is very time-consuming, tedious and subjective. In this paper, for the purpose of minimizing the annotation efforts, we propose a novel learning framework for mass detection that incorporates deep active learning (DAL) and self-paced learning (SPL) paradigm. The DAL can significantly reduce the annotation efforts by radiologists, while improves the efficiency of model training by obtaining better performance with fewer overall annotated samples. The SPL is able to alleviate the data ambiguity and yield a robust model with generalization capability in various scenarios. In detail, we first employ a few of annotated easy samples to initialize the deep learning model using Focal Loss. In order to find out the most informative samples, we propose an informativeness query algorithm to rank the large amounts of unannotated samples. Next, we propose a self-paced sampling algorithm to select a number of the most informative samples. Finally, the selected most informative samples are manually annotated by experienced radiologists, which are added into the annotated samples for the model updating. This process is looped until there are not enough most informative samples in the unannotated samples. We evaluate the proposed learning framework on 2223 digitized mammograms, which are accompanied with diagnostic reports containing weakly supervised information. The experimental results suggest that our proposed learning framework achieves superior performance over the counterparts. Moreover, our proposed learning framework dramatically reduces the requirement of the annotated samples, i.e., about 20% of all training data.