Segmentation of breast anatomy for automated whole breast ultrasound images with boundary regularized convolutional encoder–decoder network

Abstract

Abstract The segmentation of breast anatomical layers in the noisy Automated Whole Breast Ultrasound (AWBUS) images is a very challenging task. A boundary regularized deep convolutional encoder–decoder network (ConvEDNet) is proposed here to address this challenge. The training of ConvEDNet is regularized by the boundary cues, which carry geometrical constraints, with the deep supervision technique for better withstand of intrinsic speckle noise and posterior acoustic shadows in ultrasound images. The boundary regularization is denoted as deep boundary supervision (DBS) throughout this paper. The training of the ConvEDNet is further boosted with the adaptive domain transfer (ADT), which is realized with the bridge of encoder training for an edge detector on ultrasound images. Accordingly, the ADT is a two-stage of domain transfer for better landing the encoder on the ultrasound domain. The ADT can provide better network initialization than either the direct usage of pretrained model from natural images or random scratch. Based on the ADT and DBS techniques, the proposed ConvEDNet method achieves better segmentation performance compared with several classic deep learning segmentation methods on the same set of AWBUS images. The segmentation of breast anatomy may potentially assist the exclusion of false-positives for computer-aided detection to further improve the efficiency of clinical image reading.