acceptodds
Withdrawn from ICLR 2027

REPRIOR: MEASUREMENT-RECONCILED PRIOR CORRECTION FOR SPARSE-VIEW CBCT RECONSTRUCTION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sparse-view cone-beam computed tomography (CBCT) is severely under- constrained by limited X-ray projections. Learned reconstruction can provide complementary anatomical information, but using it as a prior in patient-specific Gaussian reconstruction introduces two challenges. First, the discrepancy between the learned prior and current reconstruction is spatially non-uniform: some local differences contain useful missing 3D information, whereas others reflect prior bias. Second, even a volumetrically beneficial correction can conflict with the observed X-ray measurements. Since X-ray measurements are line integrals, spa- tially separated corrections may affect overlapping rays, with their signed projec- tion responses reinforcing or partially canceling in the measurement residual, so their feasibility cannot always be assessed independently. In this paper, we intro- duce RePrior, a multi-stage progressive prior-correction framework that addresses correction actionability, measurement feasibility, and representation expressivity. Selective Correction Proposal (SCP) first identifies actionable signed local correc- tions from prior–reconstruction discrepancies. Measurement-Constrained Recon- ciliation (MCR) then jointly enforces their consistency with patient-specific X- ray measurements, while Hybrid Residual Enhancement (HRE) recovers addi- tional correction capacity through complementary voxel- and Gaussian-domain degrees of freedom around the reconciled anchor. RePrior requires neither test- time ground truth nor case-specific network fine-tuning. RePrior gains approx- imately 2.6 dB in 3D PSNR over R2-Gaussian across LUNA16, PANORAMA, and PENGWIN. Beyond prior-assisted reconstruction, it delivers additional gains in volumetric quality and observed-view projection fidelity, with consistent gains in held-out-view fidelity in the backbone-matched evaluation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Sparse-View CT Reconstruction, Medical Imaging, X-ray Computed Tomography, 3D Gaussian Splatting, Learned Volumetric Priors, Projection Consistency, Inverse Problems

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_3UoGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RePrior, a sparse-view CBCT framework that incorporates learned priors into Gaussian reconstruction through three stages: (1) Selective Correction Proposal (SCP) selects and scales local prior-derived corrections; (2) Measurement-Constrained Reconciliation (MCR) jointly reconciles these corrections with observed X-ray measurements; and (3) Hybrid Residual Enhancement (HRE) combines voxel- and Gaussian-domain residual refinement while preserving measurement consistency.

Critical strengths
  1. This paper addresses a meaningful problem in prior-assisted CBCT reconstruction: how to benefit from learned anatomical information without unconditionally transferring every prior-induced change into the final volume. Its distinction between local correction utility and patient-specific measurement compatibility provides a useful perspective on combining learned priors with Gaussian reconstruction.

  2. This paper provides a valuable discussion of the limitations of prior-guided sparse-view reconstruction. The problem analysis distinguishes spatially varying prior reliability from conflicts between volumetric improvement and projection fitting. The reported observation that 43.3% of the analyzed volumetrically beneficial candidates worsen observed-view projection error provides concrete motivation for treating these as separate problems and helps explain why direct prior supervision may be insufficient.

  3. The proposed framework connects learned correction proposals with patient-specific measurements. MCR accounts for interactions between correction responses in projection space rather than treating each local correction independently. HRE further evaluates candidate volumes using the production voxelizer and forward model before acceptance. These choices are technically well motivated and provide a controlled way to explore additional corrections under the specified measurement constraints.

  4. Although further quantitative and qualitative validation is needed to clarify the nature and reliability of the improvements, the reported results show promising gains in reconstruction quality over the evaluated baselines.

Critical weaknesses
  1. The current experiments compare RePrior mainly with projection-driven reconstruction methods such as SAX-NeRF, X-Gaussian, R²-GS, and Exact-GS. These are useful reference points, but RePrior additionally benefits from a pretrained volumetric model and a ground-truth-supervised correction predictor. It therefore remains unclear whether the reported improvements demonstrate an advantage of the proposed correction strategy over other ways of incorporating learned priors. The DeepSparse-25 baseline is helpful, but its direct prediction does not replace comparisons with prior-assisted optimization methods. Could the authors include relevant approaches such as DiffNR, DiffusionMBIR [1], and DDIP/D3IP[2] (DDIP3D) under matched acquisition conditions, while clearly reporting differences in training-data access, test-time adaptation, and computational cost? In particular, DiffNR is already discussed in Section 2.3 but is absent from the experiments.

  2. The Gaussian reconstruction VRV_R is already optimized using spatial supervision from the DeepSparse prediction VPV_P, so the two inputs are not independent estimates and may share the same errors. Defining ground-truth-derived targets for correction selection and magnitude during training does not, by itself, establish that useful corrections can be distinguished from prior bias using the observable input features at test time. In particular, Eq. (35), DA=bˉ⊙(VP−VR)D_A=\bar{b}\odot(V_P-V_R), produces zero SCP correction at locations where both inputs have the same erroneous value. Although MCR and HRE introduce additional correction freedom, the current evaluation does not sufficiently demonstrate whether these stages actually correct such shared errors. Could the authors distinguish cases where the prior contains incorrect structures from cases where both inputs miss the same structure, and provide stage-wise evidence showing how these errors are handled?

  3. It is difficult to assess the reliability of SCP’s selection mechanism from input discrepancy maps and final reconstruction images alone. At minimum, the authors should provide correction-selection maps showing which patches are accepted or rejected, together with the applied coefficient maps. These decisions should be compared post hoc with the ground-truth-derived patch-level actionability labels and the actual change in reconstruction error. Signed correction maps should separately indicate attenuation increases and decreases, since these signs do not represent acceptance and rejection. Moreover, agreement with a patch-level actionability label does not establish that every voxel benefits from the applied correction. Such an analysis would clarify whether SCP genuinely retains useful corrections and suppresses harmful ones, rather than relying on aggregate improvements to support this claim.

  4. The most apparent visual improvements in Figure 5 seem to concern artifact reduction and a cleaner local appearance. Although some boundaries appear better defined, the examples do not clearly establish whether the correction stages recover specific structures missing from the prior-assisted reconstruction, rather than primarily transferring cleaner information already present in DeepSparse. The omission of the matched Prior-R²-GS baseline from the qualitative panels further limits this assessment. Could the authors provide enlarged comparisons of the DeepSparse prior, Prior-R²-GS, intermediate stages, final reconstruction, and ground truth under identical display settings? These comparisons should examine vessels or fine bone structures that are missing or incorrectly represented in the inputs, distinguishing recovery of true anatomy from propagation or introduction of structures absent from the ground truth. The current images do not demonstrate that hallucinations occur, but they are insufficient to assess this important failure mode.

[1]. Chung, H., Ryu, D., McCann, M. T., Klasky, M. L., & Ye, J. C. (2023, June). Solving 3d inverse problems using pre-trained 2d diffusion models. In 2023 IEEE/CVF conference on computer vision and pattern recognition (CVPR) (pp. 22542-22551). IEEE.

[2]. Chung, Hyungjin, and Jong Chul Ye. "Deep diffusion image prior for efficient ood adaptation in 3d inverse problems." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

Questions for authors

Could the authors address the concerns raised in the Critical Weaknesses section with clarifications and supporting evidence?

Reviewer_Pt2uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies sparse-view cone-beam CT (CBCT) reconstruction from 25 X-ray projections. It starts from R2-Gaussian (R2-GS), a Gaussian-splatting method. R2-GS fits 3D Gaussian blobs to one patient's projections. A feed-forward network, DeepSparse-25, predicts a volume in one pass. This volume supervises R2-GS. This baseline is Prior-R2-GS.

Three stages then work on the difference between the two volumes. Selective Correction Proposal (SCP) is a small 3D network. It is trained on 155 other cases with true volumes. SCP predicts where to add the difference back, and how strongly. Measurement-Constrained Reconciliation (MCR) then refits this correction. MCR adjusts the correction region by region, and can push some regions the other way to offset others (signed compensation modes). A constraint keeps the error on the 25 observed views inside a fixed limit, the envelope. Hybrid Residual Enhancement (HRE) adds back, scaled by one number, the part of the SCP correction that MCR did not keep. HRE also adds bounded Gaussian updates under the same envelope.

On LUNA16, PANORAMA and PENGWIN (90, 90 and 50 cases), RePrior has the best 3D PSNR and SSIM. PSNR (peak signal-to-noise ratio, in dB) and SSIM (structural similarity) measure closeness to the ground truth. RePrior averages 2.63 dB above R2-GS. The core idea is sound. However, most of the gain comes from Prior-R2-GS. The reconciliation mechanisms are not isolated, and all projections are simulated.

Critical strengths
  1. The matched design makes it possible to tell where the gain comes from. The same frozen DeepSparse-25 checkpoint serves as the prior-only baseline and as the prior (page 13, lines 671 to 675). Table 2 reports vanilla, prior-assisted and full results for two backbones. Table 3 is an ablation that adds the stages one at a time on the same 60 cases. So prior quality and later corrections can be separated.

  2. The problem analysis measures a real conflict. Eq. (1) shows why corrections interact through shared rays. Figure 2(b) measures this conflict over all 230 evaluation cases. In Table 3, SCP raises 3D PSNR by 0.55 dB but lowers observed-view PSNR by 0.16 dB. MCR then loses 0.34 dB of 3D PSNR and gains 0.33 dB of observed-view PSNR.

Critical weaknesses
  1. The three proposed stages do not explain most of the performance gain, and DeepSparse-10 scores below its published numbers. The abstract claims about 2.6 dB over R2-Gaussian, plus gains beyond prior-assisted reconstruction (page 1, lines 030 to 034). R2-GS uses no learned prior, so it is the wrong baseline. The gain should be measured against the strongest baseline with a prior. Section 5.2 reports gains of 1.73, 1.93 and 1.47 dB over the best Table 1 competitor (page 7, lines 353 to 355). But Table 1 leaves out Prior-R2-GS, the strongest baseline. On LUNA16 (Table 2, 90 cases), the total gain over R2-GS is 3.46 dB, and Prior-R2-GS alone gives 2.97 dB of it. The stages add only 0.49 dB (32.09 to 32.58 dB), and 0.15 dB on the 40 held-out views. The paper gives no seed, repeat or variance information.

    In Table 3, the full method (32.43 dB) is 0.15 dB below SCP alone (32.58 dB). MCR and HRE together improve only the observed views (PSNR 46.30 to 47.10 dB). But the method fits and caps the observed-view error (Eqs. 44, 74 and 97). So this rise is not independent evidence. Held-out-view metrics exist only for LUNA16, never for the + SCP or + MCR rows.

    The paper also reports DeepSparse-10 below its published 10-view results on the same datasets (Lin et al., IEEE TMI 2026). Table 1 gives 30.74, 26.35 and 27.12 dB on LUNA16, PANORAMA and PENGWIN. The published values are 31.86, 29.42 and 29.03 dB.

    Please address these in the rebuttal:

    • (i) Report held-out-view metrics for all four Table 3 rows.
    • (ii) Add Prior-R2-GS and SCP-only rows to Table 1 for all 230 cases. Give paired case-level intervals for the stage gain.
    • (iii) In the abstract, compare against Prior-R2-GS, not R2-GS.
    • (iv) State the projection geometry, noise model and views given to DeepSparse-10. Re-run its official checkpoints on the official split.

    If SCP-only matches the full method on held-out views, the title claim is unsupported.

  2. The paper does not show that small organs and lesions survive reconstruction. The introduction says preserving local anatomical details matters, citing Lin et al. (2025) (page 1, lines 044 to 045). Section 5.2 says RePrior better follows local boundaries and fine structures (page 8, lines 408 to 411). The only evidence is boxed regions in Figures 5 and S1 to S6. Whole-volume PSNR and SSIM cannot test this. The cited work shows why. Missing a small organ or vessel may change PSNR or SSIM only in the third or fourth decimal place. Pixel metrics also correlate weakly with anatomy metrics (Pearson r of about 0.16 for SSIM and 0.30 for PSNR). A learned prior can also add anatomy that this patient does not have.

    The datasets already support this test. LUNA16 annotates lung nodules, PANORAMA annotates pancreatic lesions, and PENGWIN annotates pelvic fracture fragments. Run one frozen organ and lesion segmentation model on each reconstruction and on the ground truth. Report NSD (Normalized Surface Dice) for organs, clDice (centerline Dice) for vessels and airways, and the detection rate of the annotated lesions. Do this for RePrior, Prior-R2-GS and R2-GS. How were the boxed regions chosen? If the gain disappears on these metrics, the local-detail claim is unsupported.

  3. The reconciliation mechanisms are not tested alone, and all projections are simulated. Section 3 uses Eq. (1) to argue that corrections must be reconciled jointly (page 4, lines 208 to 211). No ablation tests this. None compares the joint solve, the compensation modes or the Gaussian branch with simpler alternatives. The only evidence is one case in Figure 4, where compensation keeps 0.21 dB of a 0.55 dB 3D gain.

    Table 3 merges regional recovery with compensation. It also merges the scalar-only HRE state (Appendix D.3) with Gaussian refinement. That state is certified for every case (page 24, line 1277), but not reported.

    Four rebuttal probes on the 60 cases would answer this.

    • (a) Count the conflicting candidates behind the premise diagnostic (page 2, lines 095 to 097), and how many MCR keeps.
    • (b) Run MCR with recovery modes only.
    • (c) Run an independent per-region accept-or-reject screen under the same envelope.
    • (d) Compare the scalar-only state with full HRE.

    Report 3D and held-out-view metrics for each. If independent screening matches the joint solve, the second contribution reduces to a data-consistency check.

    All projections are simulated. LUNA16, PANORAMA and PENGWIN contain CT volumes, not cone-beam projections. By my reading, the "measured X-ray projections" (page 2, line 076; page 13, line 687) were computed from these volumes. So Figure 2's conflicts with "X-ray measurements" are with simulated projections. The "fixed 25-view acquisition protocol" (page 13, line 676) is never defined. Real CBCT adds noise, scatter and beam hardening. If one linear TIGRE projector (page 17, line 905) both made the data and checks each correction, the constraint is easier to meet. Give the projector, geometry, angles and noise model. Add noise and a mismatched projector on the 60 cases, or test on real cone-beam projections such as the walnut collection (Der Sarkissian et al., 2019).

Questions for authors
  1. Which projector is used to check each correction? The error limits in Eqs. 74 and 75 use N, the projections rendered from the Gaussians, not the voxel projection (Eq. 38). The final recheck (page 20, lines 1070 to 1071) does not name its projector. How are the scales in Appendix D.6 fitted? No equation is given.
  2. Figures S1 and S6 carry an in-image footnote: "Existing historical-parity warning retained; metrics describe the shown reconstruction. See metric audit." What does it refer to? Do the tile metrics and Tables 1 and 3 come from one run?
  3. Which cases were used to select lambda_P, the 10K switch and the 0.05 and 0.15 dB error limits? Were any of the 230 evaluation cases among them?
  4. Which datasets supply the 155 SCP training and 20 validation cases? Do they overlap DeepSparse-25 fine-tuning? PENGWIN's 60 fine-tuning and 50 evaluation cases (page 13, lines 667 to 668) exceed its 100 public training scans (Sang et al., 2026).
Reviewer_8kEfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a method to transfer volumetric priors to Gaussian based CT reconstructions. Results are compared with INR and Gaussian based comparison methods on several popular datasets.

Critical strengths

None

Critical weaknesses
  • There has been a lot of research on the effect of priors on data fit, and specifically what kind of volumetric changes could be detected in CT data based on the specific CT geometry (e.g. analysis of null spaces and its role in prior-based reconstruction). The proposed paper does not mention this existing work, instead presenting null space-like findings as novel contributions. For example, the paragraph starting line 192 is not a novel observation, and can be explained using existing literature about null space and range space components of the CT operator.
  • The explanation of the method is lacking in clarity in the main paper. For example, V_P and V_R are mentioned around line 89 without any explanation what they represent. A bit more explanation is included later in section 3, but still very unclear 'V_P is the learned prior'. It only becomes clear what these are mathematically, and what is exactly used for them in the experiments, in the appendix. Because of this, it is difficult to properly assess the workings of the method and the experimental results. It seems that V_P is a single template volume, but many prior-based methods use generative prior distributions instead.
  • it seems the experiments are set up using a test set consisting of 40 held-out views. It would be more fair to test on held-out patients, to assess generalizability across patients.
Questions for authors
  • how does your method relate to the existing literature on analysis of null spaces of the CT operator and its effect on reconstruction quality?
  • how does your method compare with methods that use generative prior distributions?

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

Related papers

Discussion (0)

Sign in to comment.