Jointly Pre-training with Coarsely Labeled Ultrasound Images for Breast Tumor Classification

The incidence of breast cancer has been increasing in recent years, and ultrasound imaging has shown to be effective in early detection and treatment. Diagnosing breast cancer from ultrasound images with common deep learning methods usually requires a large-scale number of training data with accurate labels, which is challenging in reality. Many clinical images are labeled based on the physicians’ preliminary assessment without the biopsy verification, which leads to coarsely labeled data. Directly utilizing such a coarsely labeled clinical dataset to train a fully supervised model would not bring satisfactory accuracy. In this paper, we propose a joint self-supervised learning and coarse-label supervision learning network to overcome the limitations of the small dataset and coarse labels. Furthermore, we also propose to refine the model with a limited number of data with biopsy verified labels. We conducted extensive experiments on the collected dataset. The results show our method outperforms the state-of-the-art methods.