Source-linked AI summary
PET Quantification-Preserving Registration for Longitudinal Whole-Body PSMA PET/CT
Fryderyk Kögl, Anna Reithmeir, Daniel M. Lang, Emily Chan, Rickmer Braren, Julia A. Schnabel
TL;DR
Longitudinal PSMA PET/CT registration must align anatomy while preserving PET-derived tumor-burden measures. The method adapts coarse-to-fine LapIRN with biomarker-preservation losses, skeletal rigidity, and instance optimization; its submitted configuration improves alignment and PET preservation relative to baselines.
Problem
Longitudinal PSMA PET/CT registration must align anatomy without introducing artificial changes in PET-derived tumor-burden measures.
Source: §1 · p. 1Method
The method combines affine pre-registration, coarse-to-fine LapIRN deformation, PET biomarker-preservation losses, skeletal rigidity regularization, and test-time instance optimization.
Source: §1 · p. 2, §2.2 · p. 3, Step 2: Deformable registration · p. 3Results
81.0% DSC versus 73.3% for NiftyReg and 6.17 mm HD95 versus 7.09 mm; the submitted configuration also reduces MTV and TLG errors and non-diffeomorphic volume.
Source: §3.2 · p. 7, §4 · p. 8Takeaways & Limitations
The submitted configuration improves both alignment and PET preservation compared with baselines, supporting longitudinal analysis without artificial changes in PET-derived tumor burden.
Source: §1 · p. 1, §4 · p. 8Evaluation uses automatically generated surrogate labels, and loss weights were selected on the same 20 comparison cases, so results are exploratory rather than confirmatory.
Source: §4 · p. 8
Abstract
from arXiv · showhide
We adapt the diffeomorphic coarse-to-fine LapIRN architecture for longitudinal whole-body PSMA PET/CT registration in the Learn2Reg 2026 PSMAReg challenge. The method combines affine pre-registration, multi-modal deformable registration, and test-time instance optimization. We further introduce PET biomarker preservation losses for metabolic tumor volume (MTV) and total lesion glycolysis (TLG), including Jacobian-based constraints, and a per-structure rigidity regularization for skeletal structures. The submitted configuration achieved 81.0% DSC and 6.17 mm HD95, with MTV and TLG errors of 3.3% and 3.3%, respectively, on our evaluation protocol. Code is available at https://github.com/koegl/psmareg.
1 Introduction
Longitudinal PSMA PET/CT registration must align anatomy while avoiding artificial changes in PET-derived tumor burden. This work adapts LapIRN and adds PET biomarker preservation, skeletal rigidity, and test-time instance optimization.
1 Introduction
- Longitudinal registration supports disease-burden comparison, uptake-change localization, response assessment, dose accumulation, and quantitative analysis without artificial changes in PET-derived tumor-burden measures.The Learn2Reg 2026 PSMAReg challenge evaluates organ overlap, boundary distance, deformation regularity, and preservation of MTV and TLG.Source: §1 · p. 1
- Iterative regularized registration can achieve high accuracy but is computationally demanding for whole-body images, while learning-based methods amortize optimization into network weights.At inference, learning-based methods replace per-pair iteration with a single forward pass.Source: §1 · pp. 1–2
- The method adapts LapIRN for whole-body PSMA PET/CT and adds PET biomarker preservation losses, local bone rigidity regularization, and test-time instance optimization.LapIRN’s coarse-to-fine Laplacian pyramid is described as suited to large longitudinal deformations while maintaining smooth, invertible transformations.Source: §1 · p. 2
2 Methods
The methods combine affine pre-registration, coarse-to-fine LapIRN deformation, and optional test-time instance optimization. Training balances registration accuracy, PET biomarker preservation, and deformation regularity, with skeletal rigidity and runtime constraints shaping the final configuration.
2.1 Data and Preprocessing
- Of 358 released training cases, 146 paired cases were split patient-wise into 117 training and 29 validation cases; 20 challenge validation cases served as test data with surrogate labels.The scans were cropped to 192×192×288 voxels at 2.7344×2.7344×3.27 mm resolution; test labels were generated using TotalSegmentator and the winning autoPET III submission.Source: §2.1 · pp. 2–3
2.2 Registration Pipeline
- The pipeline applies CT-based affine pre-registration, three-level LapIRN deformable registration, then optional instance optimization of a residual stationary velocity field.LapIRN receives fixed CT/PET and affinely aligned moving CT/PET, but computes similarity on CT because post-therapy PET uptake may change; residual refinement preserves diffeomorphism and retains the best-scoring iterate.Source: §2.2 · p. 3, Step 2: Deformable registration · p. 3, Step 1: Affine pre-registration · p. 3
2.3 Training Loss Function for deformable stage
- The objective combines registration accuracy, PET quantification, and deformation regularization; PET-specific and bone terms are omitted at quarter and half resolution.CT NCC and weak CT-organ-label DSC supervise registration; MTV, Jacobian, and TLG terms preserve PET measures, while independent rigid fits regularize 61 skeletal labels and diffusion plus NDV penalties regularize deformation.Source: §2.3 · p. 4, } deformation regularization · p. 4, Preservation of PET quantification · p. 4
2.4 Submission container
- The submitted container uses predicted PET lesion masks and fast-mode CT organ labels for instance optimization, which runs until the 90-second limit.Its affine and network-predicted deformation fields are unchanged from the validation configuration.Source: §2.4 · p. 5
2.5 Model Selection
- Checkpoint and final-variant selection use the challenge composite score, since validation CT Dice continues improving after MTV and TLG errors reach their minima.The composite weights registration accuracy, biomarker preservation, and deformation regularity at 0.4, 0.4, and 0.2.Source: §2.5 · p. 5, Figure 2 · p. 6
2.6 Baselines
- The comparisons include NiftyReg, a classical iterative baseline, and ConvexAdam, the challenge’s learning-free optimization baseline.Both references and two competing methods were evaluated on the same pairs and protocol.Source: §2.6 · p. 5
2.7 Implementation Details
- Training used longitudinally motivated intensity, scaling, flipping, and field-of-view augmentations, with sequential pyramid-level training on a single H100 GPU.The three levels were trained for 60k, 60k, and 120k steps; total training time was 2d 6h 7m 6s.Source: §2.7 · p. 6
3 Results
On the 20-case surrogate-label protocol, the submitted container improved registration metrics relative to the strongest baseline while reducing PET biomarker errors and non-diffeomorphic volume. Separately reported official-server leaderboard results are for the unconstrained validation configuration.
3.1 Evaluation Protocol
- The baseline comparison evaluates three method configurations, with Ours (container) matching the hidden-test submission and Ours (no IO) isolating instance optimization.Table 1 uses the 20 official challenge validation cases with surrogate labels and a local evaluator, separate from the official leaderboard setup.Source: §3.1 · pp. 6–7
3.2 Comparison against Baselines
- 81.0% DSC versus 73.3% for NiftyReg and 6.17 mm HD95 versus 7.09 mm; the container also reduced MTV and TLG errors and non-diffeomorphic volume from 1525 to 864 ppm.Table 1 reports the 20-case surrogate-label evaluation; its cells are means over cases.Source: Table 1 · p. 7, §3.2 · p. 7
3.3 Leaderboard Performance
- The official validation leaderboard reported Ours (validation) at Dice 0.789 ± 0.0381, HD95 5.99 ± 2.39 mm, MTV error 1.56 ± 1.22%, TLG error 4.18 ± 2.49%, and NDV 0.0041 ± 0.0024%.These figures came from the official evaluation server, not the surrogate-label protocol.Source: §3.3 · p. 7
4 Discussion
The authors report that the submitted configuration improves alignment and PET preservation relative to baselines, while noting that evaluation and runtime constraints qualify the findings.
4 Discussion
- The submitted configuration improves DSC and HD95 while reducing MTV and TLG bias and non-diffeomorphic volume relative to the strongest baseline.Biomarker terms are used during training and instance optimization; per-structure rigid fits constrain skeletal structures without finite-difference stencils extending into surrounding soft tissue.Source: §4 · p. 8
- Surrogate-label evaluation and selecting loss weights on the same 20 comparison cases make the results exploratory rather than confirmatory; hidden-test results were unavailable.The rigidity term also depends on skeletal segmentation quality and is intended primarily for structures thin relative to image resolution.Source: §4 · p. 8
- The 90-second runtime constraint reduces submitted-container accuracy, largely because instance optimization requires segmentations.Future work is directed toward faster, more accurate segmentation and evaluating biomarker preservation per lesion rather than only at aggregate MTV and TLG levels.Source: §4 · p. 8