Breast Ultrasound Image Segmentation Model Based Residual Encoder

Artificial intelligence algorithms show promise for various medical imaging applications. Modern computer-aided diagnosis (CAD) systems can potentially be used for the early diagnosis of breast tumors, a leading cause of death in women worldwide. Deep learning algorithms have been applied to readily available breast ultrasound (BUS) images and provided good segmentation and classification performance. Nonetheless, this task remains challenging because US images are very noisy with a class imbalanced data distribution and inhomogenous intensity. To address these issues, we proposed a new variant of the U-Net architecture, which is commonly used for medical image segmentation. We also increased the network depth by using residual blocks (RBs). To resolve the issue of the vanishing gradient while downsampling the features, we expanded the network width by adapting a convolution path instead of a concatenation path in the original U-net. Our proposed model includes a preprocessing stage, feature extraction based on RBs, convolution-concatenation path, and a simple decoder to reconstruct the extracted features. Our proposed model showed better performance than basic U-Net and most recent models like DAL, SK-U-Net, Mobile-U-Net, and Efficient-U-Net. In particular, our model achieved a Dice coefficient and IOU of 91.5% and 84.6% on BUSIS, respectively.