The uncertainty of boundary can improve the classification accuracy of BI‐RADS 4A ultrasound image

摘要

Abstract Purpose The Breast Imaging‐Reporting and Data System (BI‐RADS) for ultrasound imaging provides a widely used reporting schema for breast imaging. Previous studies have shown that in ultrasound imaging, 90 of BI‐RADS 4A tumors are benign lesions after biopsies. Unnecessary biopsy procedures can be avoided by accurate classification of BI‐RADS 4A tumors. However, the classification task is challenging and has not been fully investigated by existing studies. For benign and malignant tumors of BI‐RADS 4A, the appearances of intra‐class tumors are highly variable, the characteristics of inter‐class tumors is overall‐similar. Discriminative features need to be found to improve classification accuracy of BI‐RADS 4A tumors. Methods In this study, we designed the network using the clinical features of BI‐RADS 4A tumors to improve the discrimination ability of network. The boundary information is embedded into the input of the network using the uncertainty. A fine‐grained data augmentation method is used to find discriminative features in tumor information embedded with boundary information. Two mathematical methods, voting‐based and variance‐based, are used to define the uncertainty of boundary, and the differences of these two definitions are compared in a classification network. Results The dataset we used to evaluate our method had 1155 2D grayscale images. Each image represented a unique BI‐RADS 4A tumor. Among them, 248 tumors were proven to be malignant by biopsy, and the remaining 907 were benign. A weakly supervised data augmentation network (WS‐DAN) was used as the backbone classification network, which showed competitive performance in finding discriminative features. Using the auxiliary input of the uncertain boundaries defined by the voting method, the area under the curve (AUC) value of our method was 0.8347 (sensitivity?=?0.7774, specificity?=?0.7459). The AUC value of the variance‐based uncertainty was 0.7789. The voting‐based uncertainty was higher than the baseline (AUC?=?0.803), which only inputs the original image. Compared with the classic classification network, our method had a significant effect improvement (p?