Abstract
Medical applications often require accurate 3D representations of complexorgans with multiple parts, such as the heart and spine. Their individual partsmust adhere to specific topological constraints to ensure proper functionality.Yet, there are very few mechanisms in the deep learning literature to achievethis goal. This paper introduces a novel approach to enforce topological constraints in3D object reconstruction using deep implicit signed distance functions. Ourmethod focuses on heart and spine reconstruction but is generalizable to otherapplications. We propose a sampling-based technique that effectively checks andenforces topological constraints between 3D shapes by evaluating signeddistances at randomly sampled points throughout the volume. We demonstrate itby refining 3D segmentations obtained from the nn-UNet architecture.