CLASSIFICATION TASK ASSISTED SEGMENTATION NETWORK FOR BREAST TUMOR SEGMENTATION IN ULTRASOUND IMAGES

Accurate and effective segmentation of breast masses plays an important role in the early stages of breast cancer treatment. However, the irregular shape of mass, the blurred mass boundary and speckle noise in breast ultrasound (BUS) images make automatic segmentation still challenging. In this paper, we propose a classification task assist (CTA) module for boosting the performance of the commonly used deep segmentation models on BUS images. This module can be easily inserted into representative segmentation models to focus on a learning task of BUS image-level classification. The incorporation of this classification learning task enable the segmentation model achieve additional supervision information to enhance the extraction of semantic context information and the localization of segmentation object. We added the CTA module to four state-of-the-art segmentation models, including codec and non-codec structures, and the experimental results show that the insertion of the CTA module into the original model results in the lowest improve-ment of 1.48% and the highest improvement of 4.29% in the intersection over union (IoU) metric.