Whole-body PET/CT registration as a representation problem

Combining frozen 3D foundation model features with GPU-accelerated iterative optimization and lesion-volume preservation wins the Learn2Reg 2026 PSMAReg Challenge with zero task-specific training.

Primary source: arXiv:2610.18804: https://arxiv.org/abs/2610.18804

What’s new: Learned deformable registration networks typically require retraining for each imaging protocol, while fast iterative solvers avoid task-specific training but struggle when optimizing raw voxel intensities alone. In Registration is a Representation Problem: Solving Whole-Body PET/CT Registration (arXiv:2610.18804, 13 Oct 2026), researchers from Massachusetts General Hospital, Harvard Medical School, and MIT CSAIL demonstrate that iterative solvers are bottlenecked by correspondence representations rather than optimization mechanics. They pair frozen multi-scale representations with the GPU-accelerated FireANTs solver in a two-stage, zero-training pipeline for longitudinal whole-body PSMA PET/CT registration.

In Stage 1, the solver estimates large-scale deformations by matching raw CT and PET intensities stacked with handcrafted MIND-SSC descriptors and 32-dimensional voxelwise features from anatomix, a 3D foundation model pretrained exclusively on synthetic volumes. Stage 2 refines small abdominal organs and tumors by optimizing generalized Dice, Chamfer, and a point-sampled region-mapping loss over coarse anatomical and lesion masks from TotalSegmentator and LION, stabilized by anatomix features. Because non-rigid warping can artificially inflate or shrink tumors and corrupt longitudinal PET response metrics, the authors construct a recall-biased lesion mask (combining organ-filtered SUV thresholding with LION predictions) and apply both Jacobian volume regularization and selective FFT-based Poisson volume projection. Without any training on the challenge cohort, the system won first place in the Learn2Reg 2026 PSMAReg Challenge, matching top CT and PET Dice scores (0.830 and 0.610) while cutting Metabolic Tumor Volume (MTV) and Total Lesion Glycolysis (TLG) biomarker errors to 1.78% and 1.39%, compared to 5.37% and 5.35% for the runner-up.

Why it matters: Longitudinal oncology imaging requires aligning whole-body scans across timepoints without introducing non-biological volume changes that distort quantitative PET therapy-response biomarkers. Coupling synthetic-pretrained 3D foundation features with fast instance optimization and explicit volume-preserving constraints delivers state-of-the-art alignment out of the box, eliminating the need to train bespoke registration networks for every clinical trial protocol.

Caveats: Fitting a single whole-body diffeomorphic field cannot properly model articulated skeletal motion or topological changes, failing in regions where limb positions change drastically between scans (such as arms crossed at baseline versus raised above the head at follow-up). Running multiple 3D feature extractors and segmentation networks alongside iterative optimization remains memory-intensive, requiring parallelized scheduling and FP16 execution to fit within 72 GB of VRAM and 79 seconds per pair on an 80 GB H100 GPU.