acceptodds
Under review as a conference paper at ICLR 2027

PRISM-NAF: Physically Realistic Intensity-Domain Spectral Modeling of Neural Attenuation Fields for Sparse-View CBCT Reconstruction

Abstract

Neural attenuation fields (NAF) provide a data-efficient approach to sparse-view cone-beam CT (CBCT) reconstruction, yet most methods retain a simplified detector model: each pixel is represented by a single center ray, while spectral and low-frequency detector effects are modeled with weakly constrained spatial degrees of freedom. Under sparse views, these mismatches can be absorbed into the reconstructed attenuation field, yielding visually plausible but systematically distorted anatomy. We propose Physically Realistic Intensity-Domain Spectral Modeling of NAF (PRISM-NAF), a detector-aware neural attenuation field that structures these degrees of freedom at the measurement level. Coherent Aperture Aggregation samples multiple physically realizable paths over each area-based detector pixel; Material-Coupled Response replaces a free voxel-wise beam-hardening spread with a compact, globally shared monotone function of attenuation; and a coarse non-negative detector-plane intensity field provides a low-bandwidth correction at the stage where additive photon flux enters the measurement. We interpret the latter as an effective nuisance correction rather than a calibrated scatter estimate. Experiments on two real anthropomorphic phantom datasets spanning three anatomies and multiple sparse-view settings show that PRISM-NAF improves perceptual reconstruction quality and vascular-structure fidelity over state-of-the-art methods, while reducing structured artifacts and preserving fine anatomical boundaries.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cone-beam CT, Sparse-view reconstruction, Neural radiance fields, Self-supervised learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2wtnOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposed a new method for sparse-view CT reconstruction through integrating rays per detector-square area, and additionally two other regularizers, a material-coupled response and a low-bandwidth correction. The paper demonstrates these on two anthropomorphic phantoms and compares with 6 other baselines, showing good performance.

Critical strengths

The paper develops a forward operator suitable for this CT projection problem and uses it for image reconstruction.

The paper compares it with existing baselines and shows good performance.

Critical weaknesses
  1. While the paper sells the "area-integration" over the rays as the primary contribution (the CAA), the ablation study in Fig 6 shows little benefit of this particular addition. The CAA seems to add the least benefit, and more benefit seems to be derived from the MCR and DPS. However, this could be made more clear with 1) error maps compared to ground truth for each ablation point and 2) metrics. Authors should add these analyses.
Questions for authors

None.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2wtn

Thank you for recognizing the forward model and the comparison with existing reconstruction methods. We appreciate your suggestion to provide error maps and quantitative metrics for each ablation configuration. We agree that the original cumulative ablation did not sufficiently distinguish CAA’s contribution from those of MCR and DPS.

[W1] Clarifying CAA’s contribution through expanded ablations

Following your suggestion, we added two center-ray controls, expanding the analysis to six configurations. The revised Fig. 6 and Table 3 now compare CAA enabled and disabled at three fixed MCR/DPS settings, with quantitative metrics and signed difference maps relative to the 720-view FDK reference after global affine intensity alignment.

CAA reduces LPIPS in all three paired comparisons on the 50-view Lungman scan, while its effects on MANIQA and vessel Dice depend on the configuration. Each entry below reports CAA off → CAA on, with the MCR/DPS settings held fixed.

MCRDPSLPIPS ↓MANIQA ↑Vessel Dice ↑
OffOff0.556 → 0.5450.341 → 0.3390.731 → 0.736
OnOff0.549 → 0.5350.361 → 0.3600.739 → 0.752
OnOn0.569 → 0.5440.357 → 0.3650.747 → 0.737

CAA’s clearest local effect is improved scan-table boundary visibility. In the rightmost column of Fig. 6, where both MCR and DPS are active, the boundary is faint with CAA disabled and more distinct with CAA enabled. This boundary lies outside both the lung ROI used for vessel Dice and the common field of view used for the difference maps. Its improvement is therefore shown directly in the reconstruction insets, separately from the reported quantitative comparisons.

The expanded analysis also clarifies the complementary effects of MCR and DPS. Reading across the lower half of Fig. 6, adding MCR to CAA makes pulmonary vessels more distinct and increases vessel Dice from 0.736 to 0.752. Adding DPS then improves the uniformity of posterior soft tissue near the spine, although the signed map shows a broad difference from FDK in this region. The radial profiles in Appendix Sec B.1, Fig. 8b explain this behavior: FDK itself exhibits cupping, so a flatter soft-tissue profile can increase local disagreement with that reference. The full configuration also achieves the highest reference-free MANIQA score in Table 3.

We have revised Sec. 5.3 to explain these distinct contributions and their metric-dependent effects more clearly. We thank the reviewer for the suggestion, which helped us make the ablation analysis more informative.

We hope that the revised manuscript and additional analyses satisfactorily address your concerns. We would be pleased to clarify any remaining points or provide further details as needed. We sincerely thank you for your careful assessment and constructive feedback, which have helped strengthen the presentation and evaluation of our work.

AuthorsOpenReview ↗

Follow-up to Reviewer 2wtn: Additional Experimental Results and Revision Summary

Dear Reviewer 2wtn,

We would be grateful if you could review our response and the expanded ablation analysis in the revised manuscript. Your suggestion led us to add two center-ray controls, yielding six configurations with quantitative metrics and FDK-referenced signed difference maps (Table 3, Fig. 6).

The revised analysis compares CAA enabled and disabled at three fixed MCR/DPS settings. CAA reduces LPIPS in all three paired comparisons, including an improvement from 0.569 to 0.544 when both MCR and DPS are active. The paired reconstructions also show a more distinct scan-table boundary with CAA. These comparisons make CAA’s contribution more explicit alongside the vessel delineation and soft-tissue uniformity effects associated with MCR and DPS.

We would appreciate your feedback on whether these additions address your concerns. If any questions or concerns remain, please let us know; we would be pleased to respond promptly and provide further clarification.

Thank you once again for devoting your time and expertise to reviewing our paper. We sincerely appreciate your constructive feedback, which has helped us strengthen both the presentation and experimental evaluation of our work.

Sincerely,
The authors of Submission 38290

Thank you! Your quantitative analysis in the rebuttal looks good to me and addressed my concerns. I will increase my score to a 4 (clear acceptance)

AuthorsOpenReview ↗

Thanks to Reviewer 2wtn

Dear Reviewer 2wtn,

Thank you very much for taking the time to carefully review our responses and revised manuscript. We sincerely appreciate your thoughtful consideration and are pleased that the expanded quantitative analysis has addressed your concerns.

Your suggestion to provide metrics and difference maps for each ablation configuration led to a more informative evaluation and helped us explain the contributions of CAA, MCR, and DPS more clearly. These additions have meaningfully strengthened our paper.

We hope that this work contributes to the development of practical self-supervised CT reconstruction methods for real measurement data and provides a useful starting point for further research.

Thank you once again for the time and care you have devoted to reviewing our work. We are sincerely grateful for your constructive feedback, which has helped us improve experimental evaluation of our paper.

Sincerely,
The authors of Submission 38290

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

This paper proposes a new CT reconstruction method PRISM-NAF which find a weakness in previous CT reconstruction methods that those methods mainly consider each pixel just has one ray and the ray is emitted from the center of pixel. However, pixel is not a point, if a method just use center ray to represent the whole pixel, that will make the physical meaning not precise and reliable. So that, for this work, they query 6 rays for each pixel. Besides that, they also introduce MCR and DPS to improve the precision phisically. In experiments, they use 2 real world datasets and compare with NAF, SAX-NeRF, and gaussian-based methods.

Critical strengths
  1. Just query a center ray for each pixel definitely is an important issue for reconstruction. The motivation in physics is correct and meaningful. And first query several rays then average the contribution is also very intuitive.

  2. Using real world dataset is pretty important which is more close to real application. Previously many methods just use synthetic datasets to evaluate their methods.

Critical weaknesses
  1. The ideas are not very novel. For example, multi-ray sampling is very normal in previous CT/computer graphics works. The reason why people use/don't use multi-ray sampling is a trade off. Multi-ray sampling has benefits in precision however it's slower than center ray sampling. I don't think this idea is novel.

  2. I think speed of reconstruction is also very important. However, you didn't post your evaluation on reconstruction speed in experiment results. In my opinions, this network looks like instant-NGP and you used multi-ray sampling, so that the speed should be slower than gaussian-based methods.

  3. Generally, previous CT reconstruction methods show the result of PSNR/SSIM/LPIPS, I feel showing PSNR/SSIM is also important. And if you can show the NVS video in sup_mat, that will be better.

Questions for authors
  1. If it's possible, could you please show the comparision of training time cost and rendering speed between your work and other works?

  2. If it's possible, could you please show the NVS video during rebuttal?

If so, and the results are good, I'm glad to increase the score.

5 comments
AuthorsOpenReview ↗

Response to Reviewer TyqW (1/2)

Thank you for recognizing the physical motivation for modeling a finite detector pixel and the value of evaluation on measured scans. In response to your comments, we clarify the contribution beyond multi-ray sampling and provide additional image-quality metrics, computational-cost comparisons, and videos.

[W1] What does the method add beyond sampling more rays? We thank the reviewer for highlighting the accuracy–cost trade-off. We agree that multi-ray sampling is established and that evaluating six paths per pixel incurs additional cost. Our contribution is the joint formulation of detector-area intensity integration with a globally shared, attenuation-dependent spectral response and a coarse detector-plane correction.

At tissue boundaries, neighboring paths can traverse different materials and contribute to one measured pixel. Even for identical paths, averaging their log projections generally differs from averaging their transmitted intensities before taking the logarithm. PRISM-NAF follows this established detector model while constraining the spectral spread through a shared function of local attenuation. At the same six paths (k=6) per pixel, adding MCR to CAA improves vessel Dice from 0.736 to 0.752 on the 50-view Lungman scan (Table 3), supporting a benefit beyond increasing the path count.

To make CAA’s contribution clearer, we expanded the ablation to compare reconstructions with and without CAA at matched MCR/DPS settings, accompanied by quantitative metrics and FDK-referenced signed difference maps (Fig. 6, Table 3). Even with MCR and DPS already active, enabling CAA makes the scan-table boundary more distinct in the reconstruction insets and lowers LPIPS from 0.569 to 0.544. These additional comparisons show that CAA contributes to boundary reconstruction beyond the effects of MCR and DPS.

[W2] Reconstruction time and rendering speed

MethodTraining (min)Peak memory (GiB)Vessel Dice ↑Rendering (ms/view)Extraction (s)
NAF (MICCAI'22 Oral)71.111.50.2521,3700.07
SAX-NeRF (CVPR'24)141.611.50.3586,0640.34
R²-Gaussian (NeurIPS'24)39.37.60.6275.61.03
Vol3DGS (CVPR'25 Highlight)75.27.60.6185.81.02
GH-NAF (CVPR'26)104.111.50.7271,8880.09
FaCT-GS (ECCV'26)3.71.90.6641.50.02
PRISM-NAF (ours, K=6)82.015.60.80310,5150.16
Training schedules of comparable duration
FaCT-GS (ECCV'26; extended, 30 min)30.410.80.7355.80.21
FaCT-GS (ECCV'26; extended, 90 min)91.112.20.7476.90.24
PRISM-NAF (ours, K=4)23.99.50.7787,2160.16

Table 7 now reports optimization time, peak GPU memory, projection rendering time, and volume extraction time for all methods. FaCT-GS is substantially faster under its standard schedule (3.7 min). To compare reconstruction quality at comparable optimization times, we extended FaCT-GS training and evaluated a reduced-cost PRISM-NAF configuration with fewer sub-rays and lower model capacity.

Our reduced-cost configuration achieves Dice 0.778 in 23.9 min, versus 0.735 in 30.4 min for FaCT-GS; our default(k=6) achieves 0.803 in 82.0 min, versus 0.747 in 91.1 min. Fig. 12 also shows clearer small vessels with PRISM-NAF. Thus, extended FaCT-GS training does not close the observed quality gap within these budgets. Gaussian methods retain a substantial full-projection rendering advantage (1.5–6.9 ms/view, versus 7.216–10.515 s/view for PRISM-NAF). Both PRISM-NAF configurations extract the reconstructed 256³ CT volume in 0.16 s. For the detailed experimental protocol and results, please refer to Appendix B.6, Fig. 12, and Table 7 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer TyqW (2/2)

[W3] PSNR/SSIM and novel-view videos

MethodPSNR: Lungman chestPSNR: Pediatric chestPSNR: Pediatric headSSIM: Lungman chestSSIM: Pediatric chestSSIM: Pediatric head
NAF (MICCAI'22 Oral)23.2523.8530.240.7550.5790.764
SAX-NeRF (CVPR'24)23.7123.7730.280.7780.5810.766
R²-Gaussian (NeurIPS'24)22.4324.3130.380.7810.5860.763
Vol3DGS (CVPR'25 Highlight)23.3722.9129.260.7900.5710.753
GH-NAF (CVPR'26)24.3123.7129.930.8120.5780.732
FaCT-GS (ECCV'26)22.7924.1730.560.7860.5820.768
PRISM-NAF24.3625.0129.710.8400.6180.751

Our initial evaluation emphasized perceptual quality and vascular preservation because our study targets artifact reduction in real CBCT measurements, while the 720-view FDK reference itself retains beam-hardening and scatter artifacts. Artifact correction can change reconstructed intensities and reduce agreement with this reference, even when tissue uniformity improves.

Nevertheless, we recognize the importance of standard reference-based quantitative comparisons. We have therefore added affine-aligned 3D PSNR and SSIM for all seven methods (Table 6), complementing the previously reported LPIPS results. We apply the same per-volume global gain-and-offset alignment to account for overall intensity-scale and offset differences across reconstructions.

PRISM-NAF achieves the highest average scores on both chest scans, while FaCT-GS scores highest on the pediatric head. The qualitative comparisons nevertheless show clear visual improvements with PRISM-NAF, particularly in soft-tissue uniformity and bone–soft-tissue boundary definition (Fig. 5). The lower head scores may partly reflect spatially varying and material-dependent intensity differences from FDK. Skull-induced beam hardening can bias enclosed soft-tissue intensities [1], while MCR and DPS improve uniformity and alter bone–soft-tissue contrast. A single global gain and offset cannot generally reconcile these changes with the cupping retained by FDK. Consequently, improved uniformity can coexist with lower reference-based scores, as discussed in Appendix B.5.

The supplementary ZIP also includes the requested novel-view projection videos and CT slice sweeps. The projection videos use the 50 common held-out Lungman views excluded from all evaluated training subsets, with quantitative projection-consistency results reported in Table 5.

Reference

[1] Cauley, Keith A., et al. "The effects of the skull on CT imaging of the brain: a skull and brain phantom study." The British Journal of Radiology 94.1119 (2021): 20200714.

We hope that the revised manuscript and additional analyses satisfactorily address your concerns. We would be pleased to clarify any remaining points or provide further details as needed. We sincerely thank you for your careful assessment and constructive feedback, which have helped strengthen the presentation and evaluation of our work.

AuthorsOpenReview ↗

Follow-up to Reviewer TyqW: Additional Experimental Results and Revision Summary

Dear Reviewer TyqW,

Thank you again for taking the time to review our manuscript and for your thoughtful, constructive feedback. We would be grateful if you could review our response and the additional results in the revised manuscript when your schedule permits.

Following your suggestions, we have added optimization time, peak GPU memory, projection rendering time, and volume extraction time, together with affine-aligned 3D PSNR/SSIM and the requested novel-view videos (Tables 6–7; supplementary material). We have also clarified our contribution beyond established multi-ray sampling and expanded the paired component ablations (Table 3, Fig. 6).

At comparable optimization durations on the 70-view Lungman scan, our reduced-cost configuration achieves Dice 0.778 in 23.9 minutes, versus 0.735 in 30.4 minutes for FaCT-GS; our default achieves 0.803 in 82.0 minutes, versus 0.747 in 91.1 minutes. The corresponding images show clearer fine vessels with PRISM-NAF (Table 7, Fig. 12).

We would appreciate your feedback on whether these additions address your concerns. If any questions or concerns remain, please let us know; we would be pleased to respond promptly and provide further clarification.

Thank you once again for devoting your time and expertise to reviewing our paper. We sincerely appreciate your constructive feedback, which has helped us strengthen both the presentation and experimental evaluation of our work.

Sincerely, The authors of Submission 38290

Reviewer_hquuOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes PRISM-NAF for sparse-view CBCT reconstruction. It integrates transmitted intensities over detector pixels, constrains spectral correction through a shared attenuation-dependent function, and adds a coarse detector-plane correction. The model is optimized from measured projections. Experiments on real anthropomorphic phantoms shows improved perceptual quality and vascular preservation compared with neural-field and Gaussian-based methods.

Critical strengths
  1. The method addresses the problem of measurement mismatch. Averaging transmitted intensities before taking the logarithm is consistent with the stated detector model.
  2. Two real phantoms acquired on different systems cover three anatomical settings and multiple view counts. Vascular metrics complement the perceptual comparisons.
  3. The appendix includes matched-representation GH-NAF [1] comparisons, held-out projections, and cupping analysis. This work also acknowledges non-identifiability and distinguishes the detector-plane correction from calibrated scatter.
Critical weaknesses
  1. The methodological novelty appears incremental. Detector-area integration follows established CT physics, the spectral formula is inherited from Park et al. [2] and GH-NAF [1], and footprint filtering follows Zip-NeRF [3]. The main contributions are only a shared attenuation-to-spread constraint and their joint integration.
  2. The evaluation does not directly establish improved attenuation accuracy. The dense-view FDK reference contains beam-hardening and scatter artifacts, while perceptual scores and Frangi-derived vessel masks provide indirect evidence. The cupping analysis is useful, but ablation studies with known attenuation and independently varied detector, spectral, and scatter effects are needed to validate the physical components.
  3. Two phantoms offer limited material diversity for test the shared monotone response assumption. Runtime and memory are also unreported, despite six sub-rays with 320 samples per path. Quantitative accuracy and cost comparisons would clarify whether the gains justify the added complexity.

[1] Oh, S. J., Lee, J. H., Lim, C. Y., Lee, D., Chung, M. J., & Kim, K. (2026). GH-NAF: Grid-Adaptive Hash-Level-Attended Neural Attenuation Fields for Discrepancy-Aware CBCT. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 41762-41772).

[2] Park, H. S., Seo, J. K., & Jeon, K. (2025). Implicit neural representation‐based method for metal‐induced beam hardening artifact reduction in X‐ray CT imaging. Medical Physics, 52(4), 2201-2211.

[3] Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., & Hedman, P. (2023, October). Zip-nerf: Anti-aliased grid-based neural radiance fields. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 19640-19648). IEEE.

Questions for authors

See strengths and weaknesses, please.

6 comments
AuthorsOpenReview ↗

Response to Reviewer hquu (1/3)

We appreciate your careful distinction between a detector model that is consistent with measurement physics and a reconstruction whose attenuation values have been independently validated. Thank you also for recognizing the value of the our evaluation and the matched GH-NAF comparison. Your comments helped us clarify both our contribution and the limits of the evidence. We respond to each point below.

[W1] What does PRISM-NAF contribute beyond CT physics, GH-NAF, and Zip-NeRF? As you point out, detector-area integration is established in CT, the spectral formula follows Park et al. [2] and GH-NAF [1], and footprint filtering builds on Zip-NeRF [3]. Our contribution is to constrain how the paths through the object and the residual signal at the detector contribute to the same measured pixel. GH-NAF predicts attenuation and discrepancy as separate outputs. When spectral spread is small, even a relative change in spread has little effect on the predicted projection (Eq. 10). MCR therefore ties local spread to attenuation through one monotone function shared across the scan (Eq. 11, Appendix A.3). CAA averages the resulting path-specific transmissions over the pixel. DPS adds a smooth, view-specific intensity correction before the logarithm.

On the 50-view scan, CAA improves boundary visibility even with MCR and DPS active. Adding MCR to CAA improves vessel Dice, and adding DPS addresses the soft-tissue shading that remains (Table 3, Figs. 6 and 8b). With GH-NAF’s encoder, decoder architecture, and footprint filtering matched to ours, the full PRISM-NAF model improves vessel Dice from 0.718 to 0.737 at 50 views and from 0.756 to 0.803 at 70 views (Table 4, Fig. 10). This comparison evaluates the complete model, since GH-NAF retains its own outputs, renderer, and losses. Giving GH-NAF our encoder, decoder architecture, and footprint filtering therefore does not close the gap. The remaining difference lies in the measurement model and training objective, which GH-NAF keeps as its own; Table 3 separates the contribution of each PRISM-NAF component.

[1] Oh, S. J., Lee, J. H., Lim, C. Y., Lee, D., Chung, M. J., & Kim, K. (2026). GH-NAF: Grid-Adaptive Hash-Level-Attended Neural Attenuation Fields for Discrepancy-Aware CBCT. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 41762-41772).

[2] Park, H. S., Seo, J. K., & Jeon, K. (2025). Implicit neural representation‐based method for metal‐induced beam hardening artifact reduction in X‐ray CT imaging. Medical Physics, 52(4), 2201-2211.

[3] Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., & Hedman, P. (2023, October). Zip-nerf: Anti-aliased grid-based neural radiance fields. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 19640-19648). IEEE.

AuthorsOpenReview ↗

Response to Reviewer hquu (2/3)

[W2/W3] Attenuation accuracy and validation of the physical components, Material diversity, runtime, and memory?

Our study aims to improve CT reconstruction from real sparse-view CBCT measurements, where detector, spectral, and scatter effects jointly influence the observed projections. For these scans, artifact-free attenuation ground truth and independently calibrated physical contributions are unavailable, limiting direct validation of each component. Controlled simulations could isolate individual effects under specified assumptions, but establishing their fidelity to these particular acquisitions would require independent system characterization that is unavailable in our study.

To strengthen the evidence on measured data, we expanded the ablation to six configurations with quantitative metrics and affine-aligned difference maps (Fig. 6, Table 3). These comparisons document the reconstruction effects of the components: improved boundary visibility with CAA, better vessel delineation with MCR, and reduced soft-tissue shading with DPS. We additionally report affine-aligned 3D PSNR and SSIM for all seven methods against the 720-view FDK reference (Table 6), extending the quantitative evaluation beyond the previously reported perceptual and vascular metrics. These scores assess agreement with the dense-view reference; independent validation of absolute attenuation accuracy remains a limitation.

Table 6. Comparison of average 3D PSNR and SSIM after affine intensity alignment to the FDK reference.

MethodPSNR: Lungman chestPSNR: Pediatric chestPSNR: Pediatric headSSIM: Lungman chestSSIM: Pediatric chestSSIM: Pediatric head
NAF (MICCAI'22 Oral)23.2523.8530.240.7550.5790.764
SAX-NeRF (CVPR'24)23.7123.7730.280.7780.5810.766
R²-Gaussian (NeurIPS'24)22.4324.3130.380.7810.5860.763
Vol3DGS (CVPR'25 Highlight)23.3722.9129.260.7900.5710.753
GH-NAF (CVPR'26)24.3123.7129.930.8120.5780.732
FaCT-GS (ECCV'26)22.7924.1730.560.7860.5820.768
PRISM-NAF24.3625.0129.710.8400.6180.751

To provide a quantitative comparison against the dense-view FDK reference, we now report affine-aligned 3D PSNR and SSIM for all methods (Table 6). PRISM-NAF performs best on both chest scans, while FaCT-GS scores highest on the pediatric head. The head results may partly reflect spatially and materially varying intensity differences: skull-induced beam hardening can bias enclosed soft-tissue intensities [4], while MCR and DPS improve uniformity and alter bone–soft-tissue contrast. A single global gain and offset cannot generally reconcile these changes with the cupping retained by FDK. Improved uniformity can therefore coexist with lower reference-based scores. We have explicitly acknowledged the resulting limitation in assessing absolute attenuation accuracy in the Limitations section. Addressing this limitation through validation against independently calibrated references remains an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer hquu (3/3)

Table 7. Computational cost and vessel Dice on the 70-view Lungman chest scan. Rendering time is reported per 758 × 758 detector projection and extraction time per 256³ attenuation volume. Bold and italic entries indicate the best and second-best vessel Dice scores, respectively.

MethodTraining (min)Peak memory (GiB)Vessel Dice ↑Rendering (ms/view)Extraction (s)
NAF (MICCAI'22 Oral)71.111.50.2521,3700.07
SAX-NeRF (CVPR'24)141.611.50.3586,0640.34
R²-Gaussian (NeurIPS'24)39.37.60.6275.61.03
Vol3DGS (CVPR'25 Highlight)75.27.60.6185.81.02
GH-NAF (CVPR'26)104.111.50.7271,8880.09
FaCT-GS (ECCV'26)3.71.90.6641.50.02
PRISM-NAF (ours, K=6)82.015.60.80310,5150.16
Training schedules of comparable duration
FaCT-GS (ECCV'26; extended, 30 min)30.410.80.7355.80.21
FaCT-GS (ECCV'26; extended, 90 min)91.112.20.7476.90.24
PRISM-NAF (ours, K=4)23.99.50.7787,2160.16

Table 7 reports computational costs across all methods. While FaCT-GS is substantially faster under its standard schedule, we further assess reconstruction quality at comparable optimization times by extending FaCT-GS training and evaluating a reduced-cost PRISM-NAF configuration with fewer sub-rays and lower model capacity, and a shorter optimization schedule alongside our default configuration. Our reduced-cost configuration achieves Dice 0.778 in 23.9 min, compared with 0.735 in 30.4 min for FaCT-GS; our default achieves 0.803 in 82.0 min, versus 0.747 in 91.1 min. The corresponding images (Fig. 12) show clearer small vessels and fewer residual artifacts with PRISM-NAF. These results demonstrate that PRISM-NAF retains a clear reconstruction-quality advantage even with reduced computation, while extended FaCT-GS training does not close the gap within the evaluated budgets.

Reference

[4] Cauley, Keith A., et al. "The effects of the skull on CT imaging of the brain: a skull and brain phantom study." The British Journal of Radiology 94.1119 (2021): 20200714.

We hope that the revised manuscript and additional analyses satisfactorily address your concerns. We would be pleased to clarify any remaining points or provide further details as needed. We sincerely thank you for your careful assessment and constructive feedback, which have helped strengthen the presentation and evaluation of our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers (summary)

Dear Area Chair and Reviewers,

Thank you for your careful assessment and constructive feedback. Our study is motivated by the practical challenge of reconstructing high-quality CT volumes directly from real sparse-view CBCT measurements. We appreciate your recognition of the physical motivation for detector-area modeling and the evaluation on real CBCT data (hquu, TyqW), as well as the comparison with six reconstruction methods (2wtn). Your comments have been valuable in helping us clarify our contributions and strengthen the experimental evaluation. We summarize the main additions below.

Contribution of each component (hquu, TyqW, 2wtn). We expanded the 50-view Lungman ablation to six configurations, with quantitative metrics and signed difference maps against FDK after global affine intensity alignment. With the other components held fixed, CAA lowers LPIPS in all three paired comparisons and makes the scan-table boundary more distinct, including when MCR and DPS are active. Adding MCR to CAA improves pulmonary vessel Dice from 0.736 to 0.752, while adding DPS reduces residual shading in homogeneous soft tissue (Table 3, Figs. 6 and 8b).

Quantitative image quality (hquu, TyqW). In response to TyqW’s request, we added affine-aligned 3D PSNR and SSIM for all seven methods. PRISM-NAF achieves the highest average scores on both chest scans, while FaCT-GS scores highest on the pediatric head (Table 6). Addressing hquu’s concern about attenuation accuracy, we clarify that these scores measure agreement with the artifact-containing FDK reference and do not establish absolute attenuation accuracy.

Computational cost and reconstruction quality (hquu, TyqW). We added per-scan optimization time, peak GPU memory, projection rendering time, and volume extraction time. On the 70-view Lungman scan, standard FaCT-GS completes optimization in 3.7 minutes. To compare reconstruction quality at comparable optimization times, we extended FaCT-GS optimization and evaluated reduced-cost and default PRISM-NAF configurations. Our reduced-cost configuration achieves vessel Dice 0.778 in 23.9 minutes, compared with 0.735 in 30.4 minutes for FaCT-GS; our default achieves 0.803 in 82.0 minutes, compared with 0.747 in 91.1 minutes. The corresponding images show clearer fine vessels with PRISM-NAF (Table 7, Fig. 12). These results demonstrate that PRISM-NAF achieves higher vascular fidelity at both evaluated budget ranges, and that extended FaCT-GS optimization does not close the observed quality gap.

Supplementary videos (TyqW). Following TyqW’s request, we added novel-view projection videos and CT slice sweeps to the supplementary ZIP. The projection videos cover 50 measured Lungman views excluded from all evaluated training subsets, following the held-out evaluation protocol described in Table 5.

We acknowledge that the limited material diversity and absence of calibrated attenuation ground truth constrain conclusions about material generalization and absolute attenuation accuracy. Broader material validation and evaluation against independently calibrated references remain important directions for future work.

We hope that the revised manuscript and additional analyses satisfactorily address your concerns. We sincerely thank the Area Chair and reviewers for their time and constructive feedback, which have helped strengthen the presentation and evaluation of our work. Please do not hesitate to let us know if you have any further questions; we would be pleased to provide additional clarification.

Sincerely,
The authors of Submission 38290

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.