Construction of a Three-dimensional Model of an Organ Using Visual SLAM and Deep Learning

Abstract

Ultrasonography is less invasive and safer than MRI or CT. However, image acquisition is dependent on the skill of the probe operator, and untrained inspectors and patients undergoing diagnosis have difficulty in understanding the three-dimensional structure of organs. For the above reasons, we propose a system to construct three-dimensional organ models from two-dimensional ultrasound images. The construction of a three-dimensional model requires ultrasound images and three-dimensional position information of the ultrasound probe. We attached a camera to the ultrasound probe and used Visual SLAM to estimate the position of the probe. The obtained ultrasound images are segmented using deep learning, and a 3D model of the organ is constructed based on the position information of the ultrasound probe. In this study, the right kidney of the phantom is targeted, and the results show that the position estimation and segmentation are highly accurate. Based on these results, a model of the right kidney was constructed.