CR‐SSL: A closely related self‐supervised learning based approach for improving breast ultrasound tumor segmentation

Abstract

AbstractBreast ultrasound (BUS) tumor segmentation can help with the early detection of breast cancer; however, the lack of human‐labeled training data is a big problem in this area. To deal with this, we propose a novel self‐supervised learning approach (CR‐SSL), where several related pretext tasks like unsupervised segmentation and edge detection are firstly learned, followed by the target tumor segmentation based fine‐tuning. Learning such related pretext tasks ensures better representation learning. Experimental study shows that CR‐SSL can improve the mean Dice and Jaccard scores by more than 4–5%, with ≥0.6 scores while having access to only 20–50 human‐labeled training samples. The best Dice scores of 0.7303 and 0.8207, and Jaccard scores of 0.7082 and 0.8015 are obtained by CR‐SSL on the test datasets. Compared with the No‐SSL baseline, CR‐SSL can achieve (10–20)% improvements in segmentation quality while working in small‐sized training dataset scenarios, suggesting its high potential practical utility.