Robust ultrasound uterus and endometrium segmentation via shape and keypoint based adversarial learning

Abstract

Segmentation of ultrasound uterus image is a challenging task due to ambiguous boundaries and heterogeneous texture in ultrasound images. In this paper, we propose new segmentation networks with shape and keypoint based adversarial learning scheme which are specialized for ultrasound uterus and endometrium image segmentation. The proposed adversarial learning scheme comes with a uterus shape discriminator and an endometrium keypoint discriminator. The uterus shape discriminator allows the segmentation network to become more robust to the ambiguous boundaries of ultrasound image by considering the shape of uterus. The endometrium keypoint discriminator determines whether the predicted endometrium regions coincide with the ground-truth endometrium keypoints which define the width and length of the endometrium. Namely, the shape discriminator is used for the shape evaluation of the segmented regions, while the keypoint discriminator is used for the evaluation of segmented shape congruency with the ground-truth keypoints. Our results on ultrasound uterus image dataset demonstrate that the proposed segmentation networks with shape and keypoint based adversarial learning scheme outperforms the conventional segmentation networks.