Wavelet-infused U-Net for Breast Ultrasound Image Segmentation

Abstract

Breast cancer is a prevalent health concern globally. The consequences of breast can- cer and its mortality can be reduced through early detection and successful treatment. However, existing safe and non-invasive methods for segmentation, such as Ultra- sound Imaging, suffer from noise, artifacts, and incomplete boundaries, which re- quire extensive training and experience for medically accurate prediction. To this end, we propose modifications to the U-Net architecture that leverage wavelet information to incorporate inter-network smoothening and multiresolution detail. This informa- tion is robust to noise and occlusions, which are prevalent in ultrasound images, and allows the model to capture subtle tissue textures and edges crucial for differentiating between lesions and unwanted artifacts. By incorporating wavelet-based sampling on the vanilla U-Net architecture, we are able to effectively almost halve the parameter count while keeping the Dice score relatively similar (74.92% Dice Score with 17M parameters vs 76.27% Dice Score with 31.1M parameters). Additionally, by incorpo- rating multiresolutional information into the decoder architecture, we are able to im- prove segmentation accuracy over the vanilla U-Net architecture (77.83% Dice Score vs 76.27% Dice Score) and IOU score (64.37% vs 62.14%). We qualitatively and quan- titatively analyze the effectiveness of incorporating multiresolution information into various other U-Net architectures (U-Net++, U-Net 3+, and Attention U-Net) and show that our methods provide more accurate segmentation results and can provide robust models that are easier to train. This can allow automated ultrasound image segmentation for more rapid and reliable cancer detection.