Nipple Segmentation and Localization Using Modified U-Net on Breast Ultrasound Images

Breast cancer causes massive deaths every year. To prevent this, the detection of a tumor in its initial stage is necessary. And correctly identifying a tumor from an ultrasound image requires years of experience and due to a few of a number of experienced specialists makes it a challenging task. In this work, a modified U-Net model named GRA U-Net is proposed to assist specialists in acceptably ascertaining a tumor in an ultrasound image. GRA U-Net is a combination of some of the existing techniques and can segment the nipple from Automated Whole Breast Ultrasound (AWBUS) images. Nipple segmentation is important as it can help in precisely locating the tumor from the outside of the breast. The segmented nipple can thus be used to locate tumor with respect to its position. There already exist so many segmentation models such as Residual-U-Net, Fcn8, Dense-U-Net and Squeeze U-Net. And on comparing them with the proposed model on parameters like accuracy, sensitivity, specificity, precision and so on. It was found that GRA U-Net delivers better performance with an accuracy of 99% and f1-measure of 92%. Thus this method could be used in bio-medical applications for improving the facilities that are present and provide a proper detection of tumor or a lesion in its initial stage.