Dual-Prior Guided Null-Space Learning with Mixture-of-Splines for Arbitrary Medical Slice Super-Resolution

A new framework for reconstructing high-resolution medical volumes from anisotropic clinical scans has been proposed, reformulating the task as a constrained recovery process to prevent anatomical hallucinations and preserve original measurements [1]. The Dual-Prior Null-Space Learning (DP-NSL) framework targets arbitrary slice super-resolution, a technique that synthesizes missing intermediate slices to build isotropic volumes from lower-resolution anisotropic acquisitions [1]. The authors argue that treating this as an unconstrained regression problem can generate anatomically implausible structures or alter the originally observed data [1]. To counter this, DP-NSL introduces two complementary priors that govern the reconstruction [1]. A Measurement-Consistent Projection (MCP) enforces a Deterministic Observation Prior, applying an exact orthogonal projection that reproduces every acquired slice with zero error and confines all learned details to the unobservable null space [1]. Within that null space, a Mixture-of-Splines (MoS) module imposes a Geometric Continuity Prior by dynamically mixing B-spline experts of different analytic orders, allowing each anatomical region to be modeled with a content-aware level of continuity [1]. The framework further incorporates a Local Spatial Consistency Decoder (LSCD) to inject local inductive bias and promote spatial coherence [1]. The MoS module replaces unconstrained multi-layer perceptron regression with spatially adaptive geometric interpolation, assigning coordinate-specific analytic continuity [2]. A B-spline of order p inherently guarantees C^{p-1} analytic continuity, and the continuous mapping is evaluated by a geometric spline expert rather than a generic neural network [2]. The authors tested DP-NSL on three CT and one MRI benchmark, reporting that it outperforms existing approaches while strictly preserving measurement consistency [1]. The paper was submitted to arXiv on 25 June 2026 and the code has been made publicly available [1].