acceptodds
Under review as a conference paper at ICLR 2027

When Uncertainty Fails to Localize Error: Voxel-Wise Reliability for Sparse-View CT Reconstruction

Abstract

Radiative Gaussian Splatting enables fast and high-quality reconstruction in sparse-view CT, but reliable localization of reconstruction errors within the imaged anatomy remains challenging. It is unclear whether current uncertainty maps can reliably indicate where the reconstruction is wrong. We investigate when uncertainty can serve as a spatial error map and present an error-localization system that predicts voxel-wise reconstruction error by combining predictive uncertainty with complementary reconstruction-derived information. Under fixed Gaussian geometry and independent rectified-Gaussian density variables, an exact closed-form predictive moment is derived in this work to enable direct computation of voxel-wise variance without repeated reconstruction sampling. The resulting uncertainty maps were evaluated over both the whole reconstruction volume and the object foreground. Experiments show that strong whole-volume uncertainty-error correspondence can conceal poor error ranking within the object foreground area. This discrepancy recurs across three tested uncertainty constructions on a 15-scene benchmark. Guided by this analysis, our system combines closed-form predictive dispersion with image, geometry, cross-method discrepancy, and projection-residual cues through a learned error predictor. The resulting system substantially improves voxel-wise error localization within the anatomy, providing a spatially resolved approach to reconstruction reliability assessment in sparse-view CT.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
closed form uncertainty, problem diagnosis, learned error predictor

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_yFyeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a novel method for estimating uncertainty for sparse-view GS-based CT reconstruction. Each Gaussian’s attenuation coefficient is modelled as an independent random variable, which leads to a closed-form uncertainty estimate due to the linear formulation of X-ray projection. Furthermore, the authors train a LightGBM error predictor on handcrafted voxel features, since the uncertainty maps correlate with error over the whole volume but insufficient within the anatomy. The predictor improves foreground error localization, driven mostly by image features and classical-reconstruction discrepancies rather than the uncertainty itself.

Critical strengths
  • Relevant finding about foreground/ background

I find that the paper convincingly shows how whole-volume evaluation inflates uncertainty–error correlation (only 0.1 within body). This is an important caution for uncertainty evaluation in CT, and it holds across three uncertainty constructions and several foreground thresholds.

  • Efficient uncertainty

The idea of making each Gaussian’s attenuation coefficient random exploits the linearity of X-ray projection. This is quite appealing, as it yields exact closed-form variance maps under the stated model, about 250× faster than Monte Carlo sampling, and the authors validate it against Monte Carlo estimates.

Critical weaknesses
  • Feature design:

The 55 voxel-wise features (window sizes, scales, directional gradients) are not justified, and no feature-importance analysis is given. It remains unclear whether the predictor is mainly an edge detector. The authors need to elaborate on this.

  • Limited and simulated evaluation:

The benchmark has only 15 scenes, several of them non-medical objects. The human experiments use simulated projections (which is ok), and there is no downstream task (e.g., lesion detection) or calibration analysis. Only ranking is assessed, not error magnitude.

  • Clarity:

The paper would benefit from clearer terminology, such as distinguishing spatial Gaussian kernels from the attenuation distributions, consistent notation and a cleaner presentation of the experimental setups, particularly the diffusion comparison and the human-anatomy U-Net fusion.

Questions for authors
  • What are the main features contributing to the final estimation?

  • Could the authors evaluate the predicted error maps on a clinically motivated downstream task, for example by testing whether high-scoring regions coincide with missed or hallucinated lesions? Could they also report calibration of the predicted error magnitudes, rather than only rank-based metrics?

6 comments
AuthorsOpenReview ↗

Response to question 1

Hi, thank you for your recognition and valuable suggestions!

For your first question, features contributing most to the final estimation are normalized reconstruction intensity and standard deviation of local intensity in a 3×3×33\times3\times3 window (intensity means the X-ray attenuation value of each voxel). Top five features and their respective contributions are listed below.

FeatureAverage split gain shareTimes entering top five
Normalized reconstruction intensity7.42%7.42\%15/1515/15
Standard deviation of local intensity in a 3×3×33\times3\times3 window6.53%6.53\%15/1515/15
Interior distance to the reconstruction foreground boundary3.67%3.67\%14/1514/15
Absolute axial offset3.55%3.55\%14/1514/15
Mean absolute Gaussian–SART difference in a 5×5×55\times5\times5 window3.31%3.31\%11/1511/15

We averaged within-model split-gain shares across 45 LightGBM models (15 leave-one-scene-out folds, three learner seeds). Fold ranks average the three seeds.

Average split gain share measures how much split gain a feature contributes when training a tree model. For example, a value of 7.42% for normalized reconstruction intensity means that, across 45 trained LightGBM models, the gain contributed by splits based on reconstruction intensity accounted for an average of 7.42% of the total split gain for each model.

Times entering top five indicates how many times, across 15 leave-one-scene-out folds, that feature’s contribution ranking fell within the top five. In each fold, contributions are recalculated.

AuthorsOpenReview ↗

Response to question 2

In terms of error magnitudes, the proposed learned error predictor will eventually give a score, which needs to be mapped to an actual error value.

Here is how we map scores to values in HU:

  1. For each voxel in the development set, prepare a training pair consisting of its predicted score and its true error.

    The ten MSD development subjects are split by subject into two groups of five. The tree model and the U-Net are trained on the 15 benchmark scenes and five of the subjects and used to score the other five. The two groups then swap roles. Each subject's scores thus come from a model that did not see that subject during training.

    The normalized ranks from the two branches are averaged within the same scan to obtain the score rr. Since reference CT images are available for the development set, the true absolute error at that voxel can be computed directly as d=∣V^−V⋆∣d = |\hat{V} - V^{\star}|. Sampling 50,000 foreground voxels per subject gives a total of 500,000 pairs of scores and errors.

  2. Use these paired data to learn a lookup table with 20 bins.

    Since intensity scales differ across scans, first compute a scale aa for each subject as the 99th percentile of absolute intensity in the foreground of that subject's reconstruction. Divide the score range from 0 to 1 into 20 equal-width bins and compute the mean d/ad/a over the voxels in each bin.

    Fit these 20 means using isotonic regression weighted by voxel count. If the estimates for two adjacent bins violate the ordering that “higher scores correspond to larger errors,” pool the bins and adjust their estimates. The resulting h(r)h(r) is the mean relative error for that score bin, learned from the development data.

  3. Look up the value for a new scan and convert it back to HU.

    For a new voxel, the model provides the score rr, and the reconstructed image provides the scale aa for that scan. The predicted absolute error is

    e^HU=4000,a,h(r).\widehat{e}_{\mathrm{HU}} = 4000\\,a\\,h(r).

    Here, aa converts the relative error back to the scan's normalized intensity scale, and 4000 then converts the intensity difference to HU because the experiment maps a CT window of width 4000 HU to [0,1][0,1]. Both test cohorts use the same lookup table. The reference CT images for the test subjects are used for the final check of the errors.

True and predicted absolute reconstruction error on 16 held-out CT scans (HU)

We applied the same frozen score-to-HU mapping to two disjoint cohorts of eight subjects each. Each cohort comprised four MSD and four LUNA16 subjects with simulated 25-view acquisitions. True is the absolute difference in HU between the reconstructed and reference CT values at a voxel. Predicted is the absolute error in HU estimated by the mapping for that same voxel.

The first column groups voxels by their fusion score rr. We used 20 equal-width score intervals to fit the mapping and 10 wider intervals to report the results. For each subject and score interval, we averaged the true errors and the predicted errors over exactly the same sampled foreground voxels. Each table entry is the median of the eight subject means in the indicated cohort, rounded to one decimal place.

Fusion scoreOriginal 8, true (HU)Original 8, predicted (HU)Additional 8, true (HU)Additional 8, predicted (HU)
[0.0, 0.1)31.327.528.124.5
[0.1, 0.2)37.136.434.932.4
[0.2, 0.3)41.443.539.838.8
[0.3, 0.4)47.750.144.544.7
[0.4, 0.5)57.357.949.551.7
[0.5, 0.6)71.167.958.460.6
[0.6, 0.7)85.283.672.474.6
[0.7, 0.8)107.9107.896.496.1
[0.8, 0.9)144.6148.4131.2132.1
[0.9, 1.0]229.7254.1232.1226.9
All sampled foreground voxels84.482.777.374.0

For example, in the highest score interval, the original cohort had a true mean absolute error of 229.7 HU and a predicted mean absolute error of 254.1 HU (subject medians). The corresponding values in the additional cohort were 232.1 HU and 226.9 HU. The final row pools all sampled foreground voxels within each subject before taking the cohort median.

To quantify calibration across the score range, we computed the absolute difference between the true and predicted bin means within each subject, then averaged these differences weighted by the number of voxels in each bin. The median of this subject-level calibration gap was 10.7 HU in the original cohort and 10.6 HU in the additional cohort.

We hope this would help clarify the calibration of predicted error magnitudes you are concerned about.

AuthorsOpenReview ↗

Reply to suggestions on the downstream task

Thank you for your valuable suggestion! Testing whether high-scoring regions coincide with missed or hallucinated lesions is a research area we haven't considered. The results will be included in our next paper, as it is already proving difficult to explain findings of this study within the conference’s page limit.

Best Wishes

Reviewer_g9x9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies when uncertainty maps fail to localize reconstruction errors in sparse-view CT. It derives closed-form uncertainty estimates for a Gaussian reconstruction model and combines them with some cues to predict voxel-wise errors. Experiments demonstrate the strong correlation between uncertainty and reconstruction errors, but reveal substantially weaker correlations within the foreground.

Critical strengths

The paper provides insight that strong whole-volume uncertainty error correlations can conceal poor error localization within the foreground. This observation highlights the importance of evaluating uncertainty maps within foreground regions.

Critical weaknesses
  1. The paper is poorly organized and difficult to follow. For example, section 3.4 introduces four types of cues for error prediction. However, the authors did not provide sufficient explanation for their selection. Furthermore, established methods(i.e., FDK SART) are also introduced without appropriate citations, which also makes the whole paper hard to follow.
  2. The comparative evaluation is difficult to assess. Although Fig.1 compares several uncertainty approaches on selected cases, the main experiments do not consolidate these comparisons across the full test set. Moreover, the evaluation primarily relies on baselines constructed within the proposed framework, leaving the advantage over established uncertainty estimation unclear. And Tab.3 shows only a negligible gain from uncertainty features, which increases the spearman correlation by 0.008 (i.e., compared IMG+DIS+RES with Full). It raises questions about their contribution to the final error-localization performance.
Questions for authors
  1. The main experiments are restricted to a 25-view setting. Although Table 10 provides results for 50- and 75-view evaluation, these use an earlier predictor configuration. The authors should explain why they conducted these experiments without using the 55-D features. And I think including diverse sparse-view settings in the main experiments would help clarify the effectiveness of the proposed method.
  2. While the paper focuses on reliability assessment for CT reconstruction, its practical utility is evaluated mainly through an error-ranking metric, i.e., Spearman correlation. Prior work has used uncertainty to improve the performance of downstream tasks[1,2]. A similar downstream evaluation would strengthen the contribution of this paper.

References: [1] Wu et.al., Active view selection with perturbed Gaussian ensemble for tomographic reconstruction. ECCV2026. [2] Jiang et.al., FisherRF: Active View Selection and Mapping with Radiance Fields using Fisher Information. ECCV2024.

3 comments
AuthorsOpenReview ↗

Reply to question 1

Hi, we are sorry about the readability. In order to comply with the page limit, we had to omit a lot of content, which could make the article hard to follow. Anyway, thank you for your patience and taking the time to read it through.

Regarding your first question, previous experiment results had showed that the view setting makes little difference to the scientific conclusion of our research. Conclusions are reproducible across different view settings. The earlier predictor configuration you mentioned was an example. It was an interim result produced in the middle of our research. In later experiments, we omitted 50- and 75- view evaluation to save time and speed up research progress. That’s why the configurations are different.

However, as you said, this might sacrifice certain level of rigor. We have conducted additional experiments that includes diverse sparse view settings based on your advice. As expected, scientific conclusions remain unchanged. The proposed method is still effective. Results are displayed below:


Median foreground Spearman correlation ρ\rho across 15 scenes. Numbers in the table are calculated by first taking the 3-seed mean of each scene, and then taking the median of the means across the 15 scenes.

Test viewsDensity SD σq\sigma_qUnified 55-D predictorEarlier 52-D predictor (just for reference)
250.1040.1040.4260.4260.3960.396
500.1000.1000.4000.4000.3740.374
750.1050.1050.4390.4390.3810.381
AuthorsOpenReview ↗

Reply to question 2

Thank you for your suggestion. We are planning to conduct downstream evaluations such as the active view selection you have mentioned, whose results will be included in our next paper, as it is already proving difficult to explain findings of this study within the conference’s page limit. In addition, we will continue to improve the paper's readability and make the experimental results more comprehensible based on your advice.

AuthorsOpenReview ↗

Give Us Your Response

Hi,

The deadline of the discussion/rebuttal phase is approaching, yet we haven't reveived any feedback from you. Our carefully crafted responses, backed by solid experimental results, deserve an equally thoughtful reply.

In terms of your rating, it seems that most of the issues come from the readability. It is difficult to present all our findings within the conference’s page limit and we have provided as detailed descriptions as possible of the experimental setup and methods. We hope you have evaluated our research findings fairly and objectively, rather than giving us a low score simply because you didn't have the patience to read it through.

Reviewer_6vwhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether voxel-wise uncertainty maps can accurately localize reconstruction errors within imaged anatomy for sparse-view CT reconstruction using Radiative Gaussian Splatting. The authors model each Gaussian's density weight as a rectified-Gaussian latent variable, max⁡(0,zi)\max(0, z_i), where zi∼N(mi,τi2)z_i \sim \mathcal{N}(m_i, \tau_i^2). After training, with the geometry held constant, they derive exact closed-form predictive mean and variance maps. These maps achieve parity with Monte Carlo estimates at an approximately 250 times lower computational cost. Evaluated on a 15-scene, 25-view benchmark, this closed-form uncertainty method achieves a whole-volume Spearman correlation with the true error of approximately 0.85. However, this correlation drops significantly to between 0.11 and 0.20 within the anatomical foreground. The authors attribute this performance gap to coarse between-region contrasts and shared errors among reconstruction members, which dispersion fundamentally cannot capture. Guided by this diagnosis, they train a LightGBM predictor using 55 distinct features, including image and geometry characteristics, dispersion, cross-method discrepancies, and projection residuals. Under leave-one-scene-out evaluation, this approach improves the foreground Spearman correlation from 0.136 to 0.399 and reduces the AUSE from 0.542 to 0.304. Moreover, a study involving 16 human subjects with simulated projections demonstrates comparable gains when the predictor is fused with a 3D U-Net.

Critical strengths
  1. The exact finite-member decomposition (Eq. 11) shows that 89-93.5% of the foreground squared error is shared across reconstruction members. Also, σq\sigma_q tracks member variation significantly better than it tracks shared error (ρ=0.427\rho = 0.427 vs. 0.1010.101). This explains why dispersion-based uncertainty struggles to rank errors within the anatomy sampling method.

  2. The evaluation framework holds the reconstruction, error, and score maps constant, changing only the region of interest. The observed drop in foreground performance persists across four thresholds, reconstruction-defined masks, and boundary erosion. The authors reinforce the trustworthiness of their conclusions by reporting scene-bootstrap intervals and transparently disclosing when grain is negligible.

  3. The closed-form moments presented in Equations 7, 8, and 9 are mathematically sound and generate the variance map in a single voxelization pass. The Monte Carlo discrepancies detailed in Table 1 decrease at the expected O(S−1/2)\mathcal{O}(S^{-1/2}) rate (variance: 42.86% →\rightarrow 15.44% →\rightarrow 7.66%), confirming that the remaining gap is attributable to sampling noise rather than a derivation error.

  4. Cross-method discrepancy () and projection-residual () features each provide additional signal beyond standard image features, aligning closely with the shared-error diagnosis. This shows that alternative reconstruction methods and raw measurements can expose structural errors that repeated iterations of a single method might uniformly replicate.

Critical weaknesses
  1. The closed -form uncertainty, positioned as the paper's headline contribution, offers minimal benefit to the final system. Removing the dispersion features only marginally reduces the foreground ρ\rho from 0.399 to 0.391, while image features alone (ρ=0.331\rho = 0.331) account for approximately 74% of the performance gain over raw dispersion. Moreover, the rectified-Gaussian moments rely on standard results from moment-propagation methods, meaning the primary methodological claims (contributions 1 and 3) do not effectively substantiate one another.

  2. The general negative conclusion regarding uncertainty maps relies on a deliberately constrained posterior. Because the geometry is fixed and the weights are mean-field, the model inherently cannot represent the shared null-space error that dominates sparse-view CT. The paper'sfiffusion results (where the posterior standard deviation reaches 0.175-0.376 compared to σq\sigma_q's 0.068 on human data) imply that the failure stems from this specific posterior design rather than a fundamental flaw in uncertainty maps. Notably, the study lacks a τ0/λreg\tau_0/\lambda_{reg} sensitivity analysis or any geometry-aware alternative.

  3. The proposed error predictor is evaluated only against its own feature subsets and raw dispersion. The benchmark lacks critical external baselines, such as a standalone U-Net error predictor, edge or gradient heuristics, a basic ∣FDK−V^∣\vert{}FDK - \hat{V}\vert{} baseline, or diffusion standard deviation. Without these comparisons, a foreground ρ\rho of 0.399 might simply encode the trivial observation that "error is high near edges". As a result, in Table 4, it is impossible to disentangle the individual contributions of the U-Net and the LightGBM tree.

  4. The evaluation is restricted to 15 heterogeneous scenes and 16 human subjects, all relying on simulated 25-view projections at a single configuration. The authors do not specify whether the simulation and reconstruction phases share the same forward operator, nor do they detail the injected noise. This omission is critical for validating the projection-residual features and leaves the method's generalization to real-world acquisitions and varying view counts entirely untested.

  5. Since the target is standardized per scene (Eq. 14), the predictor only ranks voxels within a given scene and fails to provide a calibrated absolute error magnitude. Given a modest foreground ρ≈0.40\rho \approx 0.40, coupled with the absence of error calibration or lesion-level evaluation, the paper's claim that it "substantially improves" the process and enables "reliability assessment" overstates the method's current practical utility.

Questions for authors
  1. Given that removing the dispersion features has a negligible impact on benchmark performance, what is the corresponding effect on the human cohort? If the impact on human data is similarly marginal, how do you justify framing closed-form uncertainty as a primary contribution of the paper?

  2. Does the foreground ρ\rho for σq\sigma_q change significantly under different values of τ0\tau_0 and λreg\lambda_{reg}, or when the Gaussian geometry is permitted to vary?

  3. How do standard baselines perform on the benchmark? Also, please clarify the simulation steps: did the projection and reconstruction phases use the same forward operator, and which specific noise model was applied?

4 comments
AuthorsOpenReview ↗

Reply to question 1

Thank you for your detailed comment! We will answer your questions first and then respond to the weaknesses you mentioned.

Dispersion features also have a marginal effect on human cohort performance. On human CT, we removed 6 tree features and 2 U-Net channels (all dispersion features) and retrained both branches. The results are as follows:

Ablation on Human Anatomy Numbers in Full and Ablated are medians for the two matched retrainings, with and without dispersion features. The number in Paired Δ is the median of the within-subject differences (Ablated minus Full). The 95% confidence intervals use 20,000 subject-level bootstrap resamples.

Foreground Spearman ρ (higher is better)

EvaluationFullAblatedPaired Δ95% CI
Original 8, seed 00.56590.5696-0.00007[-0.00336, +0.00195]
Original 8, three-seed mean0.54390.5425-0.00005[-0.00264, +0.00230]
Additional 8, seed 00.53710.5361-0.00082[-0.00558, +0.00062]

Normalized RMSE AUSE (lower is better)

EvaluationFullAblatedPaired Δ95% CI
Original 8, seed 00.19260.1945+0.00060[-0.00179, +0.00132]
Original 8, three-seed mean0.19880.2003+0.00066[-0.00152, +0.00160]
Additional 8, seed 00.18770.1896+0.00100[-0.00059, +0.00248]

However, we still consider the closed-form derivation as one of our major contributions for two reasons.

First, it eliminates Monte Carlo sampling error in the uncertainty estimates. With this source of error removed, we can rule out sampling noise as an explanation for the poor error localization.

Second, the derivation exploits the linear attenuation of X-rays and can be applied to other scan geometries. For a fixed Gaussian kernel gig_i and ray rr, let

hri=∫rgi(x),dx.h_{ri} = \int_r g_i(x)\\,dx.

The monoenergetic log projection pr=∑ihriaip_r = \sum_i h_{ri} a_i has variance

∑ihri2Varq(ai).\sum_i h_{ri}^{2} \mathrm{Var}_{q}(a_i).

Changing the scan geometry changes the weights hrih_{ri}. This extends the calculation to tilted rays used in industrial X-ray laminography (Chen et al., LamiGauss, 2026).

Reference

Chu Chen, Ander Biguri, Jean-Michel Morel, Raymond H. Chan, Carola-Bibiane Schönlieb, and Jizhou Li. “LamiGauss: Pitching Radiative Gaussian for Sparse-View X-Ray Laminography Reconstruction.” ICASSP 2026, pp. 11492–11496.

AuthorsOpenReview ↗

Reply to question 2

For the second question, our experiments show that foreground ρ\rho changes substantially with regularization (λreg\lambda_{\mathrm{reg}}) and position uncertainty, while varying τ0\tau_0 over the tested range has little effect. However, although increasing λreg\lambda_{\mathrm{reg}} can significantly improve foreground error ranking, it also greatly compromises the quality of reconstruction, suggesting the improvement is merely due to the fact that errors are easier to localize in reconstructions with more errors.

τ0\tau_0λreg\lambda_{\mathrm{reg}}ρown↑\rho_{\mathrm{own}}\uparrowρfixed↑\rho_{\mathrm{fixed}}\uparrowPSNR (dB)↑\text{PSNR (dB)}\uparrow
0.2510−310^{-3}0.1340.13934.74
0.5010−310^{-3}0.1360.13834.68
1.0010−310^{-3}0.1360.13834.67
0.5010−410^{-4}0.1150.14335.18
0.5010−210^{-2}0.6550.09419.65

All values are equal-weight means across the 1515 scenes. The default configuration is (τ0,λreg)=(0.5,10−3)(\tau_0,\lambda_{\mathrm{reg}})=(0.5,10^{-3}). PSNR describes the reconstruction quality. ρown\rho_{\mathrm{own}} is the foreground spearman correlation between output uncertainty under the new configuration and reconstruction error under the new configuration, while ρfixed\rho_{\mathrm{fixed}} is the foreground spearman correlation between output uncertainty under the new configuration and reconstruction error under the default configuration.

AuthorsOpenReview ↗

Reply to question 3 and some mentioned weaknesses

For the third question: How do independent baselines perform, and can edge strength explain the result?

We scored every method against the absolute error of the same Gaussian reconstruction on the same foreground voxels. The standalone 3D U-Net took only the reconstructed volume V^\hat V as input and learned to predict ∣V^−V⋆∣\lvert\hat V-V^\star\rvert. All reconstruction seeds of the test scene were held out during training. Within each scene and seed, we split the foreground at its median gradient magnitude after smoothing the reconstruction at a Gaussian scale of one voxel. We repeated the evaluation in the lower-gradient half.

Benchmark results. Results of three reconstruction seeds were averaged within each scene, then weighted the 1515 scenes equally. Higher Spearman ρ\rho and lower normalized RMSE AUSE indicate better error localization.

ScoreForeground ρ\rhoForeground AUSELow-gradient ρ\rhoLow-gradient AUSE
Density SD σq\sigma_q0.1360.1360.5420.5420.0740.0740.5590.559
Gradient magnitude0.2100.2100.4350.4350.0830.0830.5340.534
Absolute FDK–Gaussian difference0.0870.0870.5560.5560.0750.0750.5370.537
Reconstruction-only 3D U-Net0.2980.2980.3910.3910.2550.2550.4200.420
Full LightGBM0.3990.3990.3040.3040.3720.3720.3470.347

Paired comparisons. Gains are full LightGBM minus the corresponding baseline. The 9595\\% intervals use 20,00020{,}000 paired scene-bootstrap resamples.

BaselineRegionMean gain in ρ\rho9595\\% CIScenes with a gain
Gradient magnitudeLow-gradient foreground+0.289+0.289[+0.221,+0.362][+0.221,+0.362]15/1515/15
Reconstruction-only 3D U-NetAll foreground+0.101+0.101[+0.061,+0.141][+0.061,+0.141]14/1514/15

The advantage over gradient magnitude persisted even in low-gradient regions in every scene. The reconstruction-only U-Net also outperformed gradient magnitude in both regions. Its comparison with the full predictor reflects differences in inputs, supervision and architecture.

Human results for Table 4. We also evaluated the multichannel human U-Net and LightGBM separately, then averaged their voxel ranks with equal weights. All values are median foreground ρ\rho at reconstruction seed 00. Each score was evaluated against the Gaussian reconstruction error.

Human cohortLightGBMU-NetRank fusionGaussian density SDDiffusion SD
Original 880.5160.5160.5650.5650.5660.5660.0680.0680.1750.175
Additional 880.5280.5280.5240.5240.5370.5370.0450.045—

The U-Net alone nearly matched fusion on the original cohort. Fusion had the highest median on the additional cohort. The diffusion comparison was available only on the original human cohort.

Did simulation and reconstruction use the same forward operator, and what noise was added?

Simulation and reconstruction used the same cone-beam geometry and view angles. Simulation projected the reference voxel volume with TIGRE's Siddon projector. Reconstruction fitted the Gaussian parameters to those measurements using a differentiable Gaussian rasterizer, a different numerical operator. To compute residual features, we projected the reconstructed voxel volume with Siddon at the original 2525 angles and formed r=ATIGREV^−yr=A_{\mathrm{TIGRE}}\hat V-y. The residual evaluation therefore used a projector matched to the simulator.

Noise model and verification. For a clean line integral pp, let MM be the maximum over the clean training projections of that scene. We drew photon counts K∼Poisson(105e−p/M)K\sim\mathrm{Poisson}(10^5 e^{-p/M}), added Gaussian readout noise ϵ∼N(0,102)\epsilon\sim\mathcal{N}(0,10^2) in count units, and converted back to line integrals as y=max⁡0,−Mlog⁡[(K+ϵ)/105]y=\max\\{0,-M\log[(K+\epsilon)/10^5]\\}. Both human cohorts used this protocol. Noise was applied only to training views. We also replayed the benchmark generator and compared its output directly with the released data.

Benchmark checkViews checkedExact matches
Noisy training projections, all 1515 scenes375375375/375375/375
Clean held-out projections, one per scene151515/1515/15
AuthorsOpenReview ↗

Give Us Your Response

Hi,

The deadline of the discussion/rebuttal phase is approaching, yet we haven't reveived any feedback from you. Our carefully crafted responses, backed by solid experimental results, deserve an equally thoughtful reply.

In addition, your comment on our paper appears to be generated by AI. We used large language model to review our paper, and its comments are extremely similar to yours, both strengths and weaknesses. It is unsure whether your rating also comes from an AI. If so, please decalre that.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.