Whole-body PET/CT registration stops leaning on raw intensities: anatomix voxel features cut tumor-volume error to 1.78%
Synopsis
Longitudinal whole-body PET/CT registration has relied mainly on raw image intensities or sparse segmentation labels to establish correspondence; intensities are fragile under cross-modal appearance change, and labels assign only coarse identities to fine anatomy. This work instead uses per-voxel semantic representations from the 3D foundation model anatomix, combined with a hand-crafted MIND-SSC descriptor and segmentation labels, to drive the FireANTs solver through two optimization stages with lesion-volume preservation constraints. On the 2026 Learn2Reg PSMAReg Challenge test set the system ranked first without training on the challenge training set, with MTV error 1.78% and TLG error 1.39%.
Interpretation
Correspondence in whole-body PET/CT registration is now established from per-voxel semantic representations extracted by a foundation model, rather than from raw image intensities alone. Earlier registration routes either used raw pixel intensities as a similarity surrogate or relied on auxiliary segmentation labels, which assign only coarse identities to fine-grained anatomy and often cover only a subset of structures. On the 2026 Learn2Reg PSMAReg Challenge test-set leaderboard, as reported by the challenge organizers, the system ranked first with CT Dice 0.830, PET Dice 0.610, MTV error 1.78% and TLG error 1.39%.
The system stacks learned features, a hand-crafted descriptor and segmentation labels into a multi-channel input and hands it to a GPU-accelerated iterative solver that estimates deformation in two stages. Stage 1 performs greedy deformable registration on a multi-channel volume stacking CT, PET, anatomix and MIND-SSC features to remove large deformations; Stage 2 then refines organ and lesion alignment with segmentation losses. On 65 ground-truth pairs from 50 held-out training subjects, tuned FireANTs alone reached CT Dice 0.686; adding anatomix and MIND-SSC features raised it to 0.775 and cut MTV error from 15.56% to 6.31%.
Registration is constrained not to introduce additional lesion-volume change, protecting PET quantitative biomarkers. Tumor burden may genuinely change between follow-up and baseline, but the deformation itself should not add further volume change that would contaminate MTV and TLG. On the held-out training pairs, adding the volume penalty and then volume projection lowered MTV error from 6.31% to 4.95% and 3.41%, and TLG error from 6.93% to 5.36% and 4.14%, with CT Dice essentially unchanged.
Perspective
The demonstrated scope is the longitudinal whole-body PSMA PET/CT pairs of the 2026 Learn2Reg PSMAReg Challenge, downsampled by the organizers to 192×192×288 at 2.73×4.42×3.27 mm resolution, with 202 image pairs in the training set; no pretrained model was trained or fine-tuned on this dataset, which served only for hyperparameter optimization.
A single diffeomorphic model represents several motion modes across the body at once, leaving it underconstrained for articulated motion such as limbs and overconstrained by the required topological correspondence when arm position changes between time points; the authors propose bone-specific polyrigid models as future work for limb articulation.