Unsupervised Contour-Aware Registration Network for Eyeball B-mode Ocular Ultrasound Images

The study on human eyes suggests that the changes of eyeball contours are closely related to some eye diseases. For clinical trials and pathological studies, B-mode ultrasound, a noninvasive and low-cost way, contains lots of pathology information. B-mode ultrasound images have been widely used in ophthalmic diagnosis. However, few works have been done on ocular ultrasound image processing while most recent studies focus on images with high resolution, such as magnetic resonance imaging (MRI). In this work, we propose an unsupervised contour-aware registration network (CARNet) for B-mode ultrasound images to register eyeball contours. In our CARNet, a weight map is employed to emphasize the contour regions during registration. To improve registration results, we employ the gradient information to help align the contour regions, which are used to calculate the image similarity in loss function in addition to the original image. Evaluation was performed on a large variety of B-mode ultrasound image sequences, and the results demonstrate that our method works well on registration of ocular ultrasound images.