acceptodds
Withdrawn from ICLR 2027

Rethinking Real-World Thick-Slice CT Super-Resolution via Thickness-Aware Degradation Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Thick-slice CT super-resolution (SR) aims to recover high-resolution (HR) thin-slice volumes from low-resolution (LR) thick-slice scans. Existing methods primarily strengthen reconstruction priors, but commonly use geometric resampling (e.g., nearest-neighbor or linear interpolation) to model the HR-to-LR degradation. As point-wise estimators, these geometric operators determine degradation weights only from slice-center locations while ignoring physical slice thickness, thereby failing to reproduce thickness-dependent partial-volume effects and limiting the reliability of SR models in real-world settings. We formulate a slice-thickness model (STM), a slice sensitivity profile (SSP)-parameterized linear operator that explicitly separates effective slice thickness from slice spacing. The full width at half maximum (FWHM) of the SSP specifies the effective thickness, whereas the sampling locations determine the LR slice centers. STM supports both analytical and data-fitted SSPs and can be used consistently for end-to-end supervised and unsupervised iterative SR. Experiments across real-paired CT datasets and multiple SR paradigms show that STM improves degradation fidelity, SR quality, and downstream anatomical segmentation. It further converts HR-only CT volumes into scalable synthetic supervision that complements limited real pairs. These results establish physically grounded degradation modeling as a key component of real-world thick-slice CT SR.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
CT Super-Resolution, Thick-Slice CT, Medical Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6UzxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the degradation process in real-world thick-slice CT super-resolution. The authors argue that conventional geometric resampling methods overlook the effect of slice thickness and therefore cannot adequately reproduce partial-volume effects. To address this issue, they introduce the Slice-Thickness Model (STM), which uses a slice sensitivity profile (SSP) to aggregate neighbouring thin slices while separately accounting for slice thickness and spacing. The proposed operator supports analytical and data-fitted profiles and can be incorporated into both supervised SR training and iterative reconstruction. Experiments on three real-paired CT datasets demonstrate improvements in degradation fidelity and SR performance over conventional interpolation operators. The authors also report improved segmentation consistency and show that STM-generated training pairs from HR-only CT data can support supervised SR.

Critical strengths

The paper addresses a relevant problem in real-world CT reconstruction. STM is simple, interpretable, and applicable to different SR frameworks. The experimental improvements are substantial and reasonably consistent. The synthetic-data experiments provide additional practical value.

Critical weaknesses
  1. The novelty relative to existing thickness-aware degradation methods is insufficiently established. Tong (2025) already proposed a CT slice synthesis approach using normalized triangular weighting based on physical slice thickness and slice locations. The triangular SSP adopted in Eq. (7) appears mathematically equivalent to this earlier weighting scheme under corresponding parameter definitions. The physical interpretation of the SSP calibration requires further justification.
  2. The authors extend the formulation to alternative SSP profiles, data-driven kernel fitting, and different SR paradigms. However, these extensions mainly concern the parameterization and application of an existing degradation model rather than introducing a fundamentally new imaging mechanism.
  3. The claimed applicability across acquisition and reconstruction protocols is not sufficiently validated. STM uses a one-dimensional, spatially invariant SSP to approximate the formation of thick CT slices. In practice, the effective response can also depend on reconstruction kernels, scanner configurations, and acquisition geometry.
  4. The evaluation does not fully separate the benefits of the proposed framework from those of using a more realistic blur model. The paper introduces several SSP variants, including rectangular, Gaussian, triangular, and data-fitted profiles. However, their respective contributions to the final SR performance are not adequately investigated. A more detailed assessment of these components would help isolate the contributions beyond the fixed triangular aggregation model.
Questions for authors
  1. How does the triangular SSP in Eq. (7) differ mathematically from the normalized triangular weighting of Tang (2025)? If the two operators are equivalent, which specific components of STM provide the additional methodological contribution, and what existing evidence supports their individual benefits?
  2. Since the input thin slices already have finite thickness, how is their original through-plane response accounted for when setting FWHM((hT))=(T)FWHM((h_T)) = (T)? Does (T)(T) represent the width of the additional aggregation kernel or the effective thickness of the synthesized slice?
  3. Which SSP variant is used to obtain the main quantitative results in Tables 1 and 2? For the fitted SSP, which data split is used for parameter estimation? Is the same profile applied across different scanners and acquisition protocols without further adjustment?
  4. Which SSP variant is used for the main results in Tables 1 and 2? Are there any existing SR reconstruction results comparing the triangular, Gaussian, rectangular, and data-fitted SSPs under the same experimental settings? Such comparisons would help clarify whether the reported improvements mainly arise from thickness-aware aggregation or from the proposed flexible SSP parameterization.
Reviewer_cyHGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a slice-thickness model (STM) for thick-slice CT super-resolution. STM generates thick slices through normalized, slice-sensitivity-profile-weighted aggregation of thin slices, explicitly separating slice thickness from slice spacing. It supports both analytical and data-fitted profiles and can generate supervised training pairs or define data consistency in iterative reconstruction. Experiments show improved thick-slice synthesis, reconstruction quality, and segmentation consistency, alongside benefits from scaling training with synthetic pairs.

Critical strengths
  1. The experiments provide evidence that degradation modeling matters for real-world CT super-resolution. STM improves both supervised and iterative methods on real paired scans, with the reconstruction architectures kept unchanged. The gains of 2.41 and 2.69 dB for SAINT and CycleINR support the usefulness of the approach beyond a single network.

  2. STM provides a practical way to turn thin-slice CT volumes into training pairs without requiring matched thick-slice acquisitions. The CT-RATE scaling experiments show that these synthetic pairs can supplement limited real training data and improve performance on the external TCIA test set. This addresses a practical obstacle to training CT super-resolution models.

Critical weaknesses
  1. The methodological advance over Tang et al. (2025) is limited. That work already distinguishes slice thickness from spacing and uses normalized triangular weighting to synthesize thick slices, closely resembling the default STM implementation. Extending this formulation to other SSPs is useful, but the reconstruction experiments mainly compare against geometric interpolation. The SSP ablation focuses on synthesis fidelity, leaving the additional reconstruction benefit of these extensions unclear.
  2. The reconstruction experiments do not directly test a central motivation of the method: separating slice thickness from spacing. Both RPLHR-CT and TCIA use 5 mm thickness and 5 mm spacing. Although AAPM-LDCT includes overlapping slices, it is evaluated only for degradation fidelity and thickness sensitivity. SR experiments on overlapping or gapped acquisitions would better establish the practical benefit of this distinction.
  3. The downstream evaluation provides limited evidence of anatomical accuracy. Dice scores compare predictions from the same automatic segmentation model on SR and HR images, without independent manual annotations. This establishes prediction consistency but does not confirm that clinically relevant structures or lesions are faithfully preserved. A focused radiologist assessment or an annotated downstream task would strengthen the practical significance of the reported improvements.
Questions for authors

See weakness, please.

Reviewer_e1MoOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a new way to model the resolution degradation. The degradation modeling is done via a slice slice sensitivity profile (SSP). Such modeling takes different slice situations into account and is able to model standard continguous, gapped and overlapped slices. The SSP can also be learned from paired low- and high-resolution volumes. The authors show that port such degradation to existing superresolution approaches, the performance can be improved.

Critical strengths

The paper is written clearly and easy to follow.

The proposed degradation modeling is portable to existing approaches to improve their performance.

Critical weaknesses

The proposed SSP seems rather straightforward and is not substantially different than existing weighted average degradations. The SSP itself afterall is still a weighted sum, with standard version of triangle-like profile and version that can be learned from paired volumes.

The learning variant to determine the profile reamins questionable if it is applicable in practice. Modern CT machines can have all kinds of different scanning protocals, the learned profile is essentially only useful for that specific scanning conditions. The paper shows some examples on profile fitted on RPLHR-CT is applied unchanged to TCIA, but they still on the same nominal 1-mm-to-5-mm configuration. The generalization to different thicknesses, reconstruction settings, and scanner configurations remain unclear for me.

The authors describe Tang et al. (2025) as already using thickness-controlled triangular weighting and distinguishing thickness from increment. The proposed method doesn't seems to substanially different that of Tang et al. (2025) and needs to be clarified.

Overall, if the authors want to support their arguments that existing LR degradation is not good enough, they shall compare with more representative and wider range of different degradation ways, for example those combine averaging with resampling.

Questions for authors

As discussed above, the proposed SSP is essentially a different way of weighted sum. The authors also stated in the realted work that there are also slice averaging-based approaches, yet the main results section only compares the proposed SSP with centeral pixel-based interpolation degradation. Maybe the authors shall consider also compare with approaches, for example that takes points inside a pixel and average them? But then this would be just a different version of authors' proposed method with the same average weights?

It is bit questionable, why the NN, linear, cubic degraded slices on Figure 4 look very similar to each other and almost identical to high resolution one. It seems for me authors just intepolate and use that slice to compare with HR one. If NN, linear, cubic just resample, then they naturally have sharp features especially taking in-plane slices. Current Figure only confirm the proposed degradation is better than specific group of "resampling" or "interpolation" based degradations. Comparison with interpolation and averaging combined approaches is required.

Especially interesting for me is the AAPM-LDCT, as the dataset documents the resolution and overlap. However, it is not clear how the authors compare their degradation method AAPM-LDCT. If the comparison methods NN, linear, cubic only model the resolution without dealing with overlap, then the comparison to their method is again not convincing and cannot be used to support the authors' argument, that their modeling is better than exisiting ones.

The Figure 5 has similar issues. It compares the superresolution methods using proposed degradation with NN, linear, cubic. However if NN, linear, cubic is just iterpolation-based as it seems, it will for sure lose information in training data modeling that causes missing details in the results. So comparison with interpolation and weighted-average combined way of degradation is necessary. With that, the same problem is that then the degradation is basically the same as what the authors proposed.

Figure 7a shows in metrics that averaging in rectangle achieves lower LR fildelty, it would be great to show visualization as well. The way of averaging there only covers half size of the pixel there, why don't the authors compare with a averaging rectangle that is lower in height, but covers the whole pixel?

It is interesting to see the learned SSP profile looks similar to a Gaussian. Then a Gaussian or traingle would already be a good approximate in practice. Do the authors have any insights on how the profile would change with different scanning geomtries? For example how wide it will become on thick slices?

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

Related papers

Discussion (0)

Sign in to comment.