Bounded-abstaining classification for breast tumors in imbalanced ultrasound images

Abstract

Computer-aided breast ultrasound (BUS) diagnosis remains a difficult task. One of the challenges is that imbalanced BUS datasets lead to poor performance, especially with regard to low accuracy in the minority (malignant tumor) class. Missed diagnosis of malignant tumors can cause serious consequences, such as delaying treatment and increasing the risk of death. Moreover, many diagnosis methods do not consider classification reliability; thus, some classifications may have a large uncertainty. To resolve such problems, a bounded-abstaining classification model is proposed. It maximizes the area under the ROC curve (AUC) under two abstention constraints. A total of 219 (92 malignant and 127 benign) BUS images are collected from the First Affiliated Hospital of Harbin Medical University, China. The experiment tests BUS datasets of three imbalance levels, and the performance contours are analyzed. The results demonstrate that AUC-rejection curves are less affected by class imbalance than accuracy-rejection curves. Compared with the state-of-the-art, the proposed method yields a significantly larger AUC and G-mean using imbalanced BUS datasets.