Segmentation boundary guided adversarial learning for uterus landmark detection

Abstract

Uterine cancer is the second frequent cancer related to women and increases steadily. The early detection and diagnose of uterine cancer can make patient be easily recovered. Therefore, the early detection of uterine cancer is very important. For early detection of uterine cancer, experts check the length and thickness of uterus. However, due to fuzzy image quality and heterogeneous texture, the ultrasound image analysis is challenging. Especially, in the case of uterus, the shape and textures of uterus varies with menstrual cycle. Therefore, accurately diagnose with ultrasound image is too time consuming and need well trained experts. In this thesis, we propose segmentation boundary guided adversarial learning for uterus landmark detection. For the adversarial learning, the predictor predicts uterus landmark points. Then the discriminators discriminate whether given landmark points are correct or not with segmentation image. Through the adversarial learning, the landmark detection accuracy improved effectively.