Accurate Breast Tumor Identification Using Computational Ultrasound Image Features

Abstract

Breast cancer ranks the first noncutaneous malignancy incidence and mortality in women worldwide, and seriously endangers the health and life of women. Ultrasound plays a key role and yet provides an economical solution for breast cancer screening. While valuable, ultrasound is still suffered from limited specificity, and its accuracy is highly related to the clinicians, resulting in inconsistent diagnosis. To address the challenge of limited specificity and inconsistent diagnosis, in this retrospective study, we first develop a learning model based on the computational ultrasound image features and identified a set of clinically relevant features. Then, the abstract spatial interaction patterns of the ultrasound images together with the extracted features were employed for breast malignancy diagnosis. We evaluate the proposed algorithm on the Breast Ultrasound Images Dataset (BUSI). The proposed algorithm achieved a diagnostic accuracy of 89.32% and a significant area under curve (AUC) of 0.9473 with the repeated cross-validation scheme. In conclusion, our algorithm shows superior performance over the existing classical methods and can be potentially applied to breast cancer screening.