BUS-Net: A Fusion-based Lesion Segmentation Model for Breast Ultrasound (BUS) Images

Abstract

Breast cancer is the most common cancer(s) among women worldwide. The survival rate decreases if the cancer is not detected at an early stage. Breast ultrasound (BUS) is emerging as a popular modality for breast cancer detection owing to its several advantages over other modalities. We proposed a novel deep learning framework named BUS-Net for automated lesion segmentation in BUS images in this work. However, every deep learning framework has disadvantages of its own; however, the drawbacks associated with individual models can be overcome when combined. Our proposed BUS-Net is an ensemble of three popular deep learning frameworks, namely attention U-net, U-Net and SegNet. The final segmentation map generated by BUS-Net is a pixel-level fusion on the outputs of each of the individual frameworks. The potentiality of BUS-Net was tested on a publicly available dataset named BUSI dataset. This dataset consists of 647 tumor images collected from 600 different female patients. To prevent biased results, the training and test set were separate. BUS-Net framework achieved an accuracy—93.19%, precision—93.18%, recall—88.75%, dice—90.77%, and volume similarity—95.55% for lesion segmentation in the test set. The degree of correlation between the lesion region segmented by the medical experts and that segmented by BUS-Net was high ($R^2 = 0.9131$). Further, the performance of BUS-Net was also compared with the state-of-the-art techniques. This comparison showed that BUS-Net maintains a tradeoff between precision and recall, proving the robustness, efficiency, and reliability of the framework.