3D Shape Correspondence for Medical Applications Using Neural Descriptor Fields

The problem this paper is concerned with is that of unsupervised learning for point cloud representation which can be used to build anatomy correspondence without need for registration. Inspired by a recently proposed technique, called neural descriptor fields (NDFs), we derive a latent embedding for each point on a point cloud using the activation values from layers in an occupancy network. This embedding is conditioned on the point cloud and is exactly invariant to rigid-body transformations and approximately invariant to deformable transformations. We improve this technique to make it suitable for building dense correspondence, including increasing the discriminative ability of the embeddings and regularizing the latent space to enhance the anatomical plausibility after correspondence mapping. We evaluate the performance of our method using the structure of the labyrinth, the ossicles and the left lateral ventricle and we compare it to a surface registration method designed for dense anatomy correspondence. Additionally, we perform experiments to demonstrate the network’s ability to represent objects in different categories simultaneously and to process incomplete point clouds.