PMR: memory-scalable full-resolution 4D-CT deformable registration on consumer GPUs

Abstract

PMR (phase-manifold registration) is a memory-scalable implementation of 4D-CT deformable image registration: the ten breathing phases are represented as a single anchored, 2π-periodic curve, so loop closure is an algebraic identity rather than a penalty whose weight must be traded against the data term. Deformation parameters live on a coarse control grid composed with the phase weights before upsampling, which reduces the coefficient tensor by 519× (tensor-size arithmetic, not a measured allocation) and lets a full-resolution 1 mm ten-phase series complete with a measured peak of 3.25 GiB on an 8 GB consumer card. The package ships an executable verification suite that runs alongside training; it has caught fourteen documented defects in this codebase, eleven of them completely silent, with the full ledger maintained in the repository. It supports DIR-Lab and CREATIS, whose differing intensity-offset, image-origin and landmark-coverage conventions are handled explicitly and locked down by a regression test. Scope. This is about scalability, verification and reproducibility, not state-of-the-art accuracy; see LIMITATIONS.md. Ten-case mean TRE on DIR-Lab is 1.597 mm. Version 1.0.5 ships the five data-dependent checks the entry point previously listed but did not include, so a fresh clone now exits zero with six checks executed and eight explicitly skipped; it also labels derived tensor-size figures as arithmetic, and records a larger-volume memory measurement that makes the fine-grid extrapolations lower bounds. Source archived independently in Software Heritage: snapshot swh:1:snp:2871fbfe6e030ec739c82f1c5bdb3c855155fdf1. Neither dataset is redistributed here.