acceptodds
Under review as a conference paper at ICLR 2027

SUGAR: Structured Unsupervised Geometry-Aware CT Attenuation Alignment and PET Activity Reconstruction

Abstract

A misaligned computed tomography (CT)-derived attenuation map can bias time-of-flight positron emission tomography (TOF-PET) activity reconstruction. This bias can persist after subsequent image registration, motivating joint estimation of activity and CT geometry from emission data. We present SUGAR, a Structured Unsupervised Geometry-aware framework for CT attenuation Alignment and TOF-PET activity Reconstruction. SUGAR fits a differentiable acquisition model to the measured data. A coordinate network is fitted to the CT attenuation map and frozen; a global transformation and neural residual warp control its sampling coordinates. A separate positive neural field represents the PET activity. To reduce redundancy between the two geometry components, SUGAR suppresses sampled infinitesimal rigid modes in the residual deformation. It further factorizes the positive activity coefficients into a scalar magnitude and a simplex-valued composition. The activity field initially uses coarse hash-grid features, with finer features introduced on a preset schedule. In evaluations on digital brain and torso phantoms, SUGAR improves activity reconstruction over the evaluated comparison methods, while controlled tests assess attenuation-prior alignment. The results support structured geometry, activity coefficients, and resolution scheduling for coupled activity–attenuation inverse problems.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Implicit neural representations, TOF-PET reconstruction, Attenuation correction, Joint reconstruction and registration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8tfROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SUGAR, a framework that jointly reconstructs PET activity and aligns a CT-derived attenuation map using TOF-PET measurements. It addresses a practical problem: when CT and PET are misaligned, errors in attenuation correction can be absorbed into the reconstructed activity, and aligning the images afterward may not remove those errors. SUGAR represents the CT attenuation map as a frozen neural field and adjusts its sampling coordinates through a global transformation and a local deformation network. A separate neural field models PET activity. The method introduces three designs to guide optimization: removing rigid-motion components from the local deformation, separating activity coefficients into magnitude and mixture weights, and gradually introducing finer spatial features. Experiments on BrainWeb and XCAT phantoms show improved reconstruction quality over the evaluated baselines at both normal and low count levels. Additional controlled experiments examine how these designs affect geometry recovery, optimization speed, and early activity compensation. The results are promising, although validation remains largely based on simulated measurements.

Critical strengths

1.Well-motivated problem and formulation. The paper addresses a practical issue in PET reconstruction: a misaligned CT attenuation map can introduce activity errors that subsequent image registration may not remove. Jointly estimating activity and attenuation geometry through the emission model is a sensible approach.

2.Coherent optimization design. The three components address identifiable difficulties: redundancy between global and local motion, coupling between coefficient magnitude and composition, and early compensation for attenuation errors by the activity field. The full factorial ablation is useful for examining how these components interact.

Critical weaknesses

1.Evidence for reliable geometric recovery is limited. The trained gauge-correction comparison uses only one phantom and one seed. Moreover, rotation in normalized coordinates does not generally represent physical rigid motion when box side lengths differ. The deformation has no invertibility guarantee, and the PET/MR extension reports folding and out-of-box queries. These issues make it difficult to conclude that improved PET images consistently reflect correct alignment.

2.The experimental comparisons leave important uncertainties. Joint-Motion and FastMCPET undergo substantial task-specific adaptations, but their training, tuning, and computational budgets are insufficiently documented. The main evaluation uses simulated measurements with simplified acquisition physics, while the PET/MR study is not an independent test-set evaluation. The composition of the ten main reconstructions and the variability of the factorial ablation also need clarification.

Questions for authors

1.Compared with existing implicit neural representation, registration, joint reconstruction, and coarse-to-fine optimization methods, what is the indispensable new mechanism of SUGAR beyond combining these established components?

2.Beyond reducing redundancy, can the sampled rigid-mode correction provide stronger evidence for identifiability, convergence, or geometric recovery?

3.Are the CT and PET reference volumes strictly aligned before mismatch is introduced, and is the ground-truth transformation known and applied only to the attenuation prior?

4.Are the author-constructed PET sinograms, mismatch transformations, and evaluation references publicly available, together with the simulation pipeline?

5.How do the authors ensure that the author-designed simulation setup and baseline adaptations do not systematically favor SUGAR, and that all methods were tuned under comparable budgets and protocols?

1 comment
Reviewer_8tfROpenReview ↗

No Authors' Response

Since I have not found the authors' responses to my concerns, I will keep my rating.

Reviewer_YuMdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses a clinically relevant PET/CT problem. Attenuation misalignment can be absorbed into activity artifacts, and post hoc registration does not correct this misattribution. Building on MLRR, SUGAR combines a frozen CT neural field, a global-residual deformation model, and a positive activity representation within a TOF emission objective. The main framework does not require external pretraining, paired motion supervision, or PET-CT similarity losses. The explicit diagnosis of compensation between activity and geometry provides a coherent motivation for the structured design. A full factorial ablation and targeted controlled experiments examine geometry recovery, coefficient convergence, and early activity compensation.

Critical strengths

Experiments cover rigid and respiratory nonrigid mismatch under normal and low counts, with consistent mean improvements over the evaluated baselines and complementary tumor contrast recovery results.

The presentation is clear, with substantial implementation and diagnostic detail.

The authors also appropriately distinguish sampled gauge correction from global identifiability, and coefficient magnitude from total tracer activity.

Critical weaknesses

W1. The paper's main motivation is that better PET image quality alone does not prove accurate geometric alignment. However, Tables 1 and 2 only report PET reconstruction metrics, leaving the actual geometric accuracy unclear. Figure 3 only evaluates a single BrainWeb phantom under a 5∘5^\circ rotation and an optional 3 mm3\text{ mm} deformation, which does not cover the main ±30 mm\pm 30\text{ mm} / ±30∘\pm 30^\circ range or the XCAT respiratory motion cases (and the exploratory PET/MR results do not replace this). Furthermore, displacements inside homogeneous attenuation regions may not change the measured data, meaning their recovery depends heavily on implicit representation and optimization biases. To verify geometric accuracy, the main comparisons should report attenuation errors and direct geometric metrics (such as EPE where ground-truth motion is available), broken down between tissue boundaries and homogeneous regions. Also, since the model lacks explicit smoothness or invertibility constraints, the main PET/CT tables should report Jacobian determinant statistics or folding rates, similar to what is shown for PET/MR in Appendix A.6.

W2. While Table 3 shows numerical gains from each proposed component, the underlying mechanisms remain unclear and sometimes conflict with the appendix results:

  • Gauge correction: Appendix A.3.1 shows that gauge correction changes PET PSNR by at most 0.016 dB0.016\text{ dB}, yet removing it in Table 3 drops PET PSNR by 0.42 dB0.42\text{ dB} while actually improving attenuation PSNR. These numbers are hard to reconcile. In addition, because the rigid modes are defined in normalized coordinates, they do not necessarily correspond to true physical rigid motion on non-cubic domains.
  • Optimization schedule: Training starts with an activity update using a misaligned attenuation prior, and then activates all multi-scale levels at epoch 4 after just a single geometry epoch, regardless of whether alignment has converged. The 2D proxy experiment keeps geometry fixed, so it does not actually validate this schedule for joint reconstruction.
  • Polar parameterization: When the bases are learned, the scale ambiguity in polar coefficients still exists. In the fixed-basis comparison, using a single shared learning rate may unfairly favor one parameterization over the other, so each should be tuned independently.

W3. Table 1 summarizes results across "10 three-dimensional reconstructions," but it is unclear whether these 10 runs correspond to different subjects/anatomies, different motion transformations, or repeated noise realizations on the same phantom. Without knowing the exact source of variation, it is difficult to judge the statistical significance or generalizability of the improvements. Please clarify what constitutes an independent sample in these experiments, report confidence intervals for the paired differences, and account for within-anatomy correlations if multiple runs share the same underlying phantom.

Questions for authors

Seeing weaknesses.

1 comment

As there was no further author discussion, I will maintain my original rating based on the submitted response and review.

Reviewer_mtHsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors present SUGAR, a Structured Unsupervised Geometry-aware framework for CT attenuation Alignment and TOF-PET activity Reconstruction. The architecture utilizes two implicit neural representations, one for CT and one for the PET radiotracer activity. The two are linked, and the authors state that SUGAR suppresses sampled infinitesimal rigid modes in the residual deformation to reduce the redundancy between the two geometry components. In general, the method is reasonably well explained and evaluated from a PET/CT physics perspective and the results appear impressive on the phantom data. However, I see this mostly as an application of existing features without significant baselines to show how the methodological innovations succeed compared to prior reconstruction algorithms (including several dozen deep learning algorithms that are not discussed).

Critical strengths
  • Very well explained background and methodology
  • Results appear well evaluated with all relevant image reconstruction metrics
  • Addition of errors by iteration are particularly interesting.
Critical weaknesses
  • As a machine learning paper, this paper lacks sufficient baselines. While Gauge-correction, Polar-cofficients, and Continuation are ablated none of these are compared against other methods. There are dozens of machine learning papers on PET reconstruction -- only two are compared against, so its not clear how the architectural or representation learning (polar coordinates, guage correction) innovations really compare against other methods. This must be included in the text at minimum!
  • Unfortunately, I feel that even though I really enjoyed reading this paper, ICLR may not even be the right venue for this. Note that even for imaging-focused venues IEEE CVPR/WACV or journals like IEEE TCI/TRPMS comparison to other SoTA methods would be standard practice, and differences in the methods discussed at minimum.
Questions for authors
  • What are the failure modes of SUGAR?
  • What would change with real PET/CT acquisitions that is not captured by the phantom data used? How could you anticipate or mitigate this?
  • How does this solution compare to what is currently used in scanners, or in SoTA algorithms?
  • Conceptually you have a model for 3D attenuation that is provided as initialization to the PET reconstruction algorithm. For CT/attenuation alone, do you get a better fit using a NeRF ? Why is that important for the rest of the approach?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.