DP-NSL framework reconstructs medical slices with zero error consistency

Medical imaging often acquires anisotropic slices, requiring super-resolution to reconstruct isotropic volumes. Existing approaches treat this as unconstrained regression, risking anatomically implausible hallucinations or altering original data. The new DP-NSL framework reformulates the task as a constrained recovery process guided by two complementary priors. First, a Measurement-Consistent Projection (MCP) enforces a deterministic observation prior: every acquired slice is reproduced with zero error via orthogonal projection, confining all learned details to the unobservable null space. Second, a Geometric Continuity Prior is imposed by the Mixture-of-Splines (MoS) module, which dynamically mixes B-spline experts of different analytic orders to model each anatomical region with content-aware continuity. A Local Spatial Consistency Decoder (LSCD) further injects local inductive bias for spatial coherence.

Experiments on three CT datasets (likely including public benchmarks) and one MRI benchmark demonstrate that DP-NSL outperforms existing super-resolution methods while strictly preserving measurement consistency. The framework is accepted to ECCV 2026, a top computer vision conference. Code is available on GitHub. This approach is significant because it addresses a critical pain point in medical imaging: the need for high-resolution, anatomically faithful volumes without sacrificing the integrity of original clinical scans. It could improve diagnostic accuracy in CT and MRI applications, reducing false positives from hallucinated structures.