Deep Learning-based Multi-stage segmentation method using ultrasound images for breast cancer diagnosis
Se Woon Cho, Na Rae Baek, Kang Ryoung Park
Abstract
Globally, breast cancer occurs frequently in women and has the highest mortality rate. Owing to the increased need for a rapid and reliable initial diagnosis of breast cancer, several breast tumor segmentation methods based on ultrasound images have attracted research attention. Most conventional methods use a single network and demonstrate high performance by accurately classifying tumor-containing and normal image pixels. However, tests performed using normal images have revealed the occurrence of many false-positive errors. To address this limitation, this study proposes a multistage-based breast tumor segmentation technique based on the classification and segmentation of ultrasound images. In our method, a breast tumor ensemble classification network (BTEC-Net) is designed to classify whether an ultrasound image contains breast tumors or not. In the segmentation stage, a residual feature selection UNet (RFS-UNet) is used to exclusively segment images classified as abnormal by the BTEC-Net. The proposed multistage segmentation method can be adopted as a fully automated diagnosis system because it can classify images as tumor-containing or normal and effectively specify the breast tumor regions.