SSAD: Prior-driven self-supervised staircase artifact denoising via diffeomorphic flow for volumetric medical image surface reconstruction.

Abstract

Accurate 3D surface reconstruction is indispensable for morphology-sensitive tasks in volumetric medical image analysis, such as clinical diagnosis and autonomous surgery. However, the discrete nature of volumetric scans only yields voxel-represented pseudo GT surface inherently affected by staircase artifacts, severely limiting reconstruction accuracy. Current methods attempt to suppress such artifacts via approximation or smoothing but fail to address the underlying geometry and remain fundamentally bounded by the fidelity of the pseudo GT. Therefore, we propose SSAD, a prior-driven self-supervised framework which redefines reconstruction as learning a diffeomorphic flow from boundary voxels to the coherent optimal surface. It formulates boundary voxels as sparsely perturbed points, denoising discrete manifolds by hierarchically aggregating multi-scale geometry-informed features and predicting point-wise structural refinements, making the denoising process an intrinsic diffeomorphic flow of boundary voxels themselves. SSAD achieves state-of-the-art performance on multiple benchmarks, with a 2.15%-22.27% improvement across various accuracy metrics compared to the top-tier counterparts. It coherently generalizes to unseen anatomical instances in a zero-shot manner, with an average increase of 4.02% in chamfer distance. Furthermore, SSAD demonstrates a super-resolution capability, with the accuracy of its reconstructed surfaces comparable to that achieved by voxel data at 2-5× the input resolution. Our code is available at https://github.com/SlimeChosen/SSAD.