acceptodds
Under review as a conference paper at ICLR 2027

KB-Splatting: Differentiable Kaiser-Bessel Splatting for Real X-ray Projection-Enabled CT Reconstruction

Abstract

Sparse-view cone-beam computed tomography (CBCT) can reduce harmful X-ray patient exposure, but recovering an accurate three-dimensional attenuation field from limited X-ray projections remains severely ill-posed. To address this, recent differentiable CT reconstruction methods have adopted anisotropic Gaussian primitives. Yet CT volumes contain sharp attenuation transitions, whereas Gaussian profiles are smooth and couple spatial extent with kernel decay. Accurately representing these transitions can therefore require several small primitives, driven by adaptive control heuristics. Thus increasing memory and optimization cost. These methods are also evaluated mainly on synthetic digitally reconstructed radiographs (DRRs), so their performance on physical clinical scanners remains unclear. Consequently, we introduce KB-Splatting, a differentiable CT reconstruction framework based on anisotropic generalized Kaiser-Bessel primitives. Our KB primitives have exact compact support per kernel and independently control support size and profile concentration, decoupling where attenuation is represented from how it varies within that tissue region. Their analytic cone-beam projection also allows us to optimize these kernel parameters directly from real X-ray projections. We also introduce a simple reconstruction-guided population control scheme that uses volumetric residuals and local parent kernel selection to add and relocate primitives in poorly reconstructed regions. We evaluate KB-Splatting on established DRR benchmarks and introduce a paired multi-system CBCT dataset containing real measured projections, documented acquisition geometries with coordinate-convention checks, and reference CT volumes from one commercial CT system and one experimental laboratory system. Experiments compare KB-Splatting with neural-field and Gaussian-splatting baselines, where it surpasses current state-of-the-art (SOTA) methods in attenuation PSNR, with a mean gain of 0.93 dB at 2x training speed. Finally, we extend KB-Splatting to RGB novel-view synthesis, testing the same representation under both additive X-ray formation and occlusion-aware rendering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Computed Tomography, 3D Reconstruction, Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mhZLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents KB-Splatting, a differentiable reconstruction method for sparse-view CBCT. Its main design choice is to replace Gaussian primitives with anisotropic generalized Kaiser-Bessel kernels: compactly supported, with kernel extent and profile sharpness exposed as separate parameters, which the authors motivate by the sharp attenuation edges typical of CT. Since the kernels admit an analytic cone-beam projection, the model is optimised against measured projections rather than a surrogate. Primitives are inserted and relocated according to the local reconstruction error. The submission also introduces a paired CBCT dataset with real projections from two scanners, and reports a mean gain of 0.93 dB in attenuation PSNR over state-of-the-art methods, with training roughly 2× faster.

Critical strengths

The choice of basis is well argued. Kaiser-Bessel kernels are compactly supported and have a closed-form X-ray transform, and the authors expose support width and profile sharpness as two independent parameters — at the sharp attenuation edges of CT that is the relevant degree of freedom, and it is what a Gaussian cannot give, since a single σ fixes both at once. Because the substitution happens inside an otherwise unchanged 3DGS pipeline, other splatting-based CT methods should be able to take it up without re-deriving their optimiser.

The paired dataset may turn out to be the most reusable part of the submission. Two systems contribute measured projections together with reference volumes, so the evaluation runs on real acquisition; synthetic DRRs, as the authors point out, carry neither scatter nor calibration error.

The reported numbers are 0.93 dB PSNR and roughly 2× faster training. If those hold under a matched tuning budget, this is accuracy that is not being paid for in compute — a combination rarer than it should be.

Critical weaknesses

The quantitative evidence is thin. Everything in the paper comes from 15 scans on two scanners, and only as averages — no per-scanner breakdown, no spread across repeats. Fig. 4 does not carry the aggregate either: on the neck case R²-Gaussian is ahead (44.54 vs. 43.34 dB). A 0.93 dB mean over 15 cases would be a good deal more convincing with a per-case table and some estimate of variance.

Calling the setting sparse-view is a stretch. The sparsest configuration still uses about 150 projections (the other two use roughly 300 and 600), and going from 600 down to 150 costs KB-Splatting only 0.24 dB. That curve is essentially flat, which suggests the experiments never enter the ill-posed regime the paper is motivated by. Results in the few-dozen-views range are what this claim needs.

The ablation does not isolate the proposed primitive. In the controlled comparison of Table 2, KB kernels improve on a matched Gaussian by 0.04–0.12 dB and run 8% faster; the headline numbers in Table 1 are 0.93 dB and about 2× over R²-Gaussian. The gap between those two settings is the issue: most of the improvement would then have to come from population control, initialisation or implementation detail rather than from the basis itself — which is the paper's central claim.

Two things around the baselines and the reference volumes need a direct answer. The neural baselines land near 15 dB, some 30 dB below every splatting method; that pattern usually means they were not tuned for real data rather than that they are 30 dB worse. And the reference volumes are themselves reconstructions, as the authors acknowledge, so differences of a fraction of a dB may fall inside the reference's own error. Neither point is fatal, but both are cheap for the authors to clarify in a rebuttal.

Questions for authors

see above

3 comments
AuthorsOpenReview ↗

Response to Reviewer mhZL (1/3)

Thank you for acknowledging our argument for a more expressive primitive in volumetric CT reconstruction. We also appreciate your thoughtful recognition of the value of our dataset. Our goal is to provide a rendering pipeline and a dataset that will progress the field.

Q1. The quantitative evidence is thin. Everything in the paper comes from 15 scans on two scanners, and only as averages — no per-scanner breakdown, no spread across repeats. Fig. 4 does not carry the aggregate either: on the neck case R$^2$-Gaussian is ahead (44.54 vs. 43.34 dB). A 0.93 dB mean over 15 cases would be a good deal more convincing with a per-case table and some estimate of variance.

We agree that aggregate scores alone do not show how reconstruction quality varies across scans and scanners. Following your suggestion, Tables 6–11 now report results for each anatomy and inanimate object, grouped by dataset and acquisition system, alongside the corresponding baseline results. This makes case-specific outcomes explicit, including the original Neck result in which R$^2$-Gaussian achieved higher PSNR. For our expanded ablation in Table 2, we report means and sample standard deviations over three random seeds, showing both the average differences between configurations and their variability across repeated runs. The table is reproduced above for your convenience.

Q2. Calling the setting sparse-view is a stretch. The sparsest configuration still uses about 150 projections (the other two use roughly 300 and 600), and going from 600 down to 150 costs KB-Splatting only 0.24 dB. That curve is essentially flat, which suggests the experiments never enter the ill-posed regime the paper is motivated by. Results in the few-dozen-views range are what this claim needs.

We agree that the original 50% and 25% settings do not fully test reconstruction in the few-dozen-views regime. We have therefore extended training and evaluation to 10% (27–80 projections) and 5% (14–40 projections) of the available views. Tables 6–11 now give the exact projection count for each scan, since the number of acquired views differs across datasets. All methods use the same selected projections for each comparison. These additional settings test clinically sparser acquisitions and show how reconstruction quality changes as fewer views are available. We report per-case results and provide corresponding reconstructions and error maps in Figures 7–22.

AuthorsOpenReview ↗

Response to Reviewer mhZL (2/3)

Q3. The ablation does not isolate the proposed primitive. In the controlled comparison of Table 2, KB kernels improve on a matched Gaussian by 0.04–0.12 dB and run 8% faster; the headline numbers in Table 1 are 0.93 dB and about 2×2\times over R$^2$-Gaussian. The gap between those two settings is the issue: most of the improvement would then have to come from population control, initialisation or implementation detail rather than from the basis itself — which is the paper's central claim.

We agree that the matched primitive comparison and the comparison against R$^2$-Gaussian measure different effects. The Gaussian control in Table 2 tests a change of primitive within a common reconstruction framework, whereas R$^2$-Gaussian is a complete method with its own rendering, optimization, and population-control implementation. The larger improvement over R$^2$-Gaussian therefore cannot be attributed entirely to replacing the Gaussian primitive.

Following your suggestion, we expanded the ablation to 12 Loop-X scans at 50% and 25% views and three random seeds. The matched study compares Gaussian, fixed- and learned-concentration truncated Gaussian, and fixed- and learned-β\beta KB primitives. For the primitive-only comparisons, we hold projection data, losses, population budgets, and training schedules fixed. We also retain an ordinary learned-β\beta KB control alongside the joint density–scale–concentration update to measure the optimizer's contribution separately.

To test reconstruction guidance, we compare selected-view FDK initialization with guided allocation, the same initializer with unguided allocation, and random initialization with unguided allocation. The last condition uses no reconstructed volume for initialization or placement. We further compare physical per-ray integration along each detector pixel's source-to-detector ray with the baseline center-ray approximation, keeping the remaining settings fixed within each primitive family. These controls separate the effects of the primitive, concentration learning, initialization, continued guidance, and rendering approximation.

Tables 4–5 report the paired attenuation-PSNR difference between learned-β\beta KB and the matched Gaussian control as 1.42±0.0061.42\pm 0.006 dB, together with their training times and variability across seeds. The expanded ablation table is reproduced above for your convenience.

AuthorsOpenReview ↗

Response to Reviewer mhZL (3/3)

Q4. Two things around the baselines and the reference volumes need a direct answer. The neural baselines land near 15 dB, some 30 dB below every splatting method; that pattern usually means they were not tuned for real data rather than that they are 30 dB worse. And the reference volumes are themselves reconstructions, as the authors acknowledge, so differences of a fraction of a dB may fall inside the reference's own error. Neither point is fatal, but both are cheap for the authors to clarify in a rebuttal.

We thank you for identifying this discrepancy. After an extensive audit of all neural methods, we identified inconsistent PSNR peak conventions: the neural baseline evaluator uses a reference-percentile intensity range, whereas the splatting evaluator uses a fixed attenuation peak. We corrected this mismatch by recomputing neural PSNR from the saved attenuation RMSE under the same fixed-peak convention. The updated LoopX neural averages are approximately 41–43 dB, eliminating the apparent 30 dB gap. We updated all quantitative and qualitative results accordingly.

Regarding the CT references, the reported differences measure agreement with reconstructed reference volumes. Although the reference volumes are reconstructions rather than independent attenuation ground truth, the accompanying acquisition metadata, measured scan geometry, and noise profile make them well-characterized estimates of the underlying attenuation structure. They therefore provide a strong practical basis for evaluating reconstruction fidelity, while retaining some uncertainty from acquisition noise and reconstruction artifacts.

Reviewer_922mOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KB-Splatting, a differentiable reconstruction framework for sparse-view cone-beam CT using anisotropic generalized Kaiser–Bessel (KB) primitives. The authors argue that conventional Gaussian primitives are limited by their infinite support and the coupling between spatial coverage and profile decay, which can make sharp attenuation transitions difficult to represent efficiently.

To address these limitations, the proposed KB primitives provide exact compact support and independently parameterize spatial extent and profile concentration. The authors derive analytical cone-beam line integrals and implement a differentiable GPU rasterizer. They further introduce a reconstruction-guided population control mechanism that uses residuals relative to an FDK reconstruction to guide primitive allocation and relocation.

The method is evaluated on 15 measured CBCT scans from two scanner systems, using different projection fractions. The reported results show an average attenuation PSNR improvement of 0.93 dB over R2-Gaussian, together with substantially reduced training time. Additional experiments investigate the primitive representation, rendering approximations, and transfer to RGB novel-view synthesis.

Critical strengths

1. A meaningful exploration of non-Gaussian primitives for differentiable CT reconstruction.

The paper presents a coherent motivation for reconsidering Gaussian primitives in sparse-view CT reconstruction. Unlike Gaussian kernels, generalized Kaiser–Bessel primitives have exact compact support and allow the spatial support and interior profile concentration to be parameterized separately.

The authors derive the corresponding cone-beam projection formulation and implement an efficient differentiable rasterization pipeline. This constitutes a technically meaningful adaptation of classical tomographic basis functions to modern explicit primitive optimization.

The matched Gaussian control is particularly appreciated because it attempts to distinguish improvements due to the representation from those due to implementation choices.

2. Evaluation on real measured X-ray projections with promising accuracy–efficiency trade-offs.

A significant strength is the evaluation on measured projections from two different CBCT systems rather than relying exclusively on synthetic digitally reconstructed radiographs.

The authors also document acquisition geometries and address coordinate-convention issues, which are practically important for reproducible tomographic reconstruction.

The reported improvements over R2-Gaussian in reconstruction accuracy and training time are encouraging. Even though the individual contributions require further investigation, the complete framework appears practically useful under the evaluated settings.

Critical weaknesses

1. Insufficient evidence that the proposed Kaiser–Bessel primitive is necessary or that its specific properties explain the reported improvements.

The central motivation of the paper is that Gaussian primitives are poorly suited to representing sharp attenuation transitions because their spatial extent and decay profile are coupled. The proposed KB primitives are intended to address this limitation through compact support and independent concentration control.

However, it is not clear that these properties uniquely motivate the use of Kaiser–Bessel functions, rather than a more general modification of the Gaussian representation.

In particular, finite support can also be introduced through truncated Gaussian kernels, potentially with an independently parameterized truncation radius. Other compactly supported radial basis functions or generalized interpolation kernels could provide similar flexibility. The paper does not sufficiently establish why the particular KB profile should be preferred over these simpler or alternative representations.

Although Section 3.3 mentions truncated-Gaussian controls, the reported matched representation ablation primarily compares Gaussian, fixed-concentration KB, and learnable-concentration KB primitives. This does not fully disentangle the benefits of compact support, independent profile control, and the specific functional form of the KB basis.

Moreover, the controlled ablation in Table 2 shows a relatively small improvement in attenuation PSNR, from 51.413 dB for matched Gaussian to 51.531 dB for learnable KB, corresponding to only 0.118 dB. In contrast, the full pipeline reports an average improvement of approximately 0.93 dB over R2-Gaussian.

These experiments are conducted under different configurations and should not be directly compared as a numerical decomposition. Nevertheless, the difference raises an important question about how much of the overall improvement is attributable to the kernel representation itself, as opposed to reconstruction-guided population control, rendering approximations, or other implementation differences.

The population control mechanism also relies on FDK-based residual sampling, which introduces an additional structural prior and potentially changes where the model allocates representation capacity. The current ablations primarily investigate variations of the residual sampling policy rather than fully separating the contribution of population control from that of the primitive family.

This is consequential because the primary methodological contribution is the claim that KB primitives provide a more suitable representation for sparse-view CT. While the results demonstrate that the proposed pipeline is effective, they do not yet convincingly establish that the specific KB representation is responsible for the reported advantages or that it is preferable to simpler compact-support alternatives.

2. An important ambiguity in reference-volume usage, together with limitations in the evaluation protocol.

A particularly concerning issue is the inconsistent description of how reference CT volumes are used during reconstruction.

Section 3.5 states that reference CT volumes are used only for qualitative evaluation and that the FDK initializer used for residual seeding is reconstructed from the training-view projections. The reproducibility statement similarly indicates that reference volumes are used only for evaluation.

However, Appendix F.3 states that the reference volumes are used for evaluation and for the coarse residual proposal field described in Section 3. This appears inconsistent with the earlier description.

The distinction is important because the reconstruction-guided population control in Eq. (5) uses a volumetric residual to determine where new primitives should be allocated. Even if reference volumes never enter the projection loss directly, using them for residual-guided primitive placement would introduce additional information into the reconstruction process. In sparse-view settings, this could potentially include information not available from the selected training projections.

The manuscript therefore needs to unambiguously establish whether the residual field is computed exclusively from an FDK reconstruction of the selected training views or whether the reported reference volumes are involved.

Beyond this ambiguity, the evaluation references are themselves reconstructed volumes produced by proprietary or analytical algorithms rather than independent attenuation ground truth. As the authors acknowledge, agreement with these references does not necessarily establish improved physical attenuation accuracy or diagnostic quality.

Finally, the end-to-end speed advantage is difficult to attribute specifically to the KB representation. The matched primitive ablation shows nearly identical training times for learnable KB and Gaussian primitives, while the full benchmark reports a substantially larger speed advantage over R2-Gaussian. This suggests that implementation and pipeline differences may contribute significantly to the reported efficiency improvement.

These concerns materially affect the reliability and interpretation of the main experimental claims. In particular, the reference-volume ambiguity must be resolved before the sparse-view reconstruction results can be assessed with confidence.

Questions for authors

1. Reference-volume usage and possible information leakage (essential clarification).

Section 3.5 states that the FDK initializer used for residual seeding is reconstructed only from the training-view projections, and the reproducibility statement indicates that reference volumes are used exclusively for evaluation.

However, Appendix F.3 states that reference volumes are used for evaluation and for the coarse residual proposal field.

Could the authors clarify exactly which volume is used to compute the residual in Eq. (5), particularly for the 25% and 50% projection settings?

Is the residual field always computed using an FDK reconstruction obtained exclusively from the selected training projections, without access to the reference CT or excluded projections?

If the reference volume is involved in population allocation, could the authors provide results using only information available from the training measurements?

2. Necessity of Kaiser–Bessel primitives (additional evidence).

The main motivation is that KB primitives offer compact support and independent control of spatial extent and concentration.

However, similar properties could potentially be achieved using truncated Gaussian kernels with independent truncation radii or other compactly supported radial basis functions.

What specific properties of the Kaiser–Bessel family make it more suitable for sparse-view CT reconstruction than these alternatives?

A controlled comparison with a truncated Gaussian whose support radius and interior decay can be adjusted independently would be particularly informative.

Additionally, could the authors clarify the motivation for selecting KB order (m=1) as the final reconstruction configuration, given that different orders exhibit different boundary continuity properties?

3. Attribution of performance improvements (additional evidence).

The matched representation ablation in Table 2 reports a relatively small improvement from replacing Gaussian primitives with learnable KB primitives, whereas the complete framework achieves a substantially larger improvement over R2-Gaussian.

Could the authors provide a controlled component-wise ablation separating the effects of the primitive family, learnable concentration, reconstruction-guided population control, and rendering implementation?

In particular, how does a Gaussian-based implementation perform when equipped with the same reconstruction-guided population control and optimization settings as KB-Splatting?

This comparison would help establish whether the main improvements arise from the proposed representation or from the broader optimization pipeline.

4. Reconstruction accuracy and efficiency under matched evaluation conditions (clarification).

The reported reference volumes are generated using analytical or proprietary reconstruction procedures rather than independent attenuation ground truth.

How sensitive are the reported improvements to the evaluation reference and its reconstruction artifacts?

Additionally, could the authors clarify whether the approximately twofold training speed advantage is measured under fully matched hardware, primitive budgets, and implementation conditions? Reporting final primitive counts and support-overlap statistics would help connect the claimed compact-support advantage to the observed runtime improvements.

Post-review comments I appreciate the authors' detailed response and the additional experiments.

The clarification regarding reference-volume usage addresses my concern about potential information leakage. The authors confirm that the FDK initializer is reconstructed exclusively from the selected training projections and that the reference CT volumes are used only for evaluation. This resolves the ambiguity in the original manuscript.

The expanded ablation study also substantially strengthens the justification for using KB primitives. In particular, the comparisons with fixed and learnable truncated Gaussians directly address my question about the necessity of the proposed representation. The results show that support truncation alone cannot explain the improvements achieved by KB primitives. Furthermore, KB maintains its advantage without FDK-based reconstruction guidance and under a reduced primitive budget. These experiments provide stronger evidence that the primitive family itself contributes to the reconstruction improvements.

The additional analysis of contributing primitives per detector pixel also supports the claimed computational advantage. The reduced number of contributors provides a plausible explanation for the lower rendering cost.

Nevertheless, I have two remaining concerns.

First, the original matched representation ablation reported a 0.118 dB improvement over Gaussian primitives with comparable training times. The revised study reports substantially larger improvements in both reconstruction quality and training efficiency. I understand that the authors identified a confounding issue in the previous Gaussian control and introduced a separate Gaussian implementation. However, the revised results would benefit from a more explicit explanation of the differences between the two experimental settings and the validation of the corrected Gaussian renderer. This is particularly important because the new results provide the main evidence for the advantage of KB primitives.

Second, the reported number of experiments needs clarification. The stated 24 configurations, 12 scans, two projection fractions, and three random seeds correspond to 1,728 runs rather than the reported 1,584. Please clarify whether some configurations were excluded.

I also note that learning the concentration parameter provides only a modest additional PSNR improvement over fixed-concentration KB in the expanded ablation. Thus, the results provide stronger evidence for the effectiveness of the KB primitive family than for the necessity of learning the concentration parameter.

Overall, the rebuttal addresses my main concerns. The expanded experiments provide meaningful evidence that the reported improvements are not solely attributable to support truncation or reconstruction-guided population allocation. Although the revised Gaussian comparison needs further clarification, the additional results substantially strengthen the methodological justification and empirical validation.

I therefore revise my rating from 2 to 3.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 922m (1/4)

We are grateful for your recognition of the technical contribution of adapting Kaiser–Bessel functions to differentiable cone-beam reconstruction and the value of the matched Gaussian control. We also appreciate your recognition of the reconstruction accuracy and training efficiency of our method on real X-ray data from multiple CBCT systems.

Q1. Reference-volume usage and possible information leakage (essential clarification). Section 3.5 states that the FDK initializer used for residual seeding is reconstructed only from the training-view projections, and the reproducibility statement indicates that reference volumes are used exclusively for evaluation. However, Appendix F.3 states that reference volumes are used for evaluation and for the coarse residual proposal field. Could the authors clarify exactly which volume is used to compute the residual in Eq. (5), particularly for the 25% and 50% projection settings? Is the residual field always computed using an FDK reconstruction obtained exclusively from the selected training projections, without access to the reference CT or excluded projections? If the reference volume is involved in population allocation, could the authors provide results using only information available from the training measurements?

We agree that Appendix F.3 does not sufficiently distinguish the evaluation reference from the training-view FDK initializer and Coarse Residual Field (CRD) pipeline. This distinction is essential for assessing the accuracy of our sparse-view CT reconstructions. As such, we will replace the first paragraph with:

The evaluation references are scanner-reconstructed DICOM volumes for the Loop-X and CTorch FBP reconstructions for the experimental CT machine. We use these volumes strictly for evaluation and do not use them for primitive initialization, residual-guided placement, or pipeline optimization. Furthermore, for each sparse-view training subset, we construct a separate FDK initializer using only the selected X-ray projections. CRD then compares this sparse-view FDK reconstruction with the current primitive field on a coarse 32332^3 grid, converting their residual into sampling probabilities that guide primitive placement. This residual serves solely as a placement proposal and is neither minimized as a training loss nor differentiated during optimization. Consequently, neither projections outside the selected training subset nor voxel values from the evaluation reference enter the CT reconstruction process of our KBS pipeline.

We use CRD to seed where in the attenuation field new and relocated KB primitives are placed. We then optimize primitive parameters by minimizing mean squared error and SSIM loss between rendered and measured training views, with additional penalties on primitive density and scale. As such, CRD provides an allocation prior without introducing additional projection measurements or reference-volume supervision.

AuthorsOpenReview ↗

Response to Reviewer 922m (2/4)

Q2. Necessity of Kaiser–Bessel primitives (additional evidence). The main motivation is that KB primitives offer compact support and independent control of spatial extent and concentration. However, similar properties could potentially be achieved using truncated Gaussian kernels with independent truncation radii or other compactly supported radial basis functions. What specific properties of the Kaiser–Bessel family make it more suitable for sparse-view CT reconstruction than these alternatives? A controlled comparison with a truncated Gaussian whose support radius and interior decay can be adjusted independently would be particularly informative. Additionally, could the authors clarify the motivation for selecting KB order (m=1) as the final reconstruction configuration, given that different orders exhibit different boundary continuity properties?

Thank you for your valuable feedback. A fair and rigorous comparison is our priority. As suggested, we expanded the ablation to compare Gaussian, fixed and learned truncated Gaussian (TG), and fixed and learned-β\beta KB across 12 LoopX cases with three seeds. Fixed TG uses a fixed truncation cutoff, whereas learned TG optimizes a per-primitive cutoff. Both primitives optimize basis position, scale, rotation, and density. Fixed-β\beta KB holds concentration constant, while learned-β\beta KB uses joint density–support–concentration updates. We evaluate these variants under full/guided allocation, unguided allocation, no reconstruction prior, and per-ray rendering.

Table 2 reports the results at 25% views. Under the full/guided configuration, learned-β\beta KB improves mean attenuation PSNR over Gaussian by 1.42 dB and reduces LPIPS from 0.075 to 0.056. Gaussian truncation improves PSNR by at most 0.05 dB, while fixed-β\beta KB already achieves 46.81 dB, compared with 46.86 dB for learned-β\beta KB with joint update. These comparisons suggest that support truncation alone does not account for the observed advantage. Furthermore, halving KB's average primitive population from 225k to 113k retains a 1.26 dB advantage over the full-budget Gaussian control. The Gaussian primitive family and KB retain the same optimization settings. As such, these results attribute the metric improvements to the KB primitive. Figure 5 complements Table 2 with matched right-hand and phantom-neck crops, showing clearer bone boundaries and better overall rendering fidelity for KB variants. These enlarged attenuation regions make local reconstruction differences visible alongside the quantitative results.

Furthermore, in our experiments, m=1m=1 provides the best balance between reconstruction quality and training cost. It is the lowest integer order for which the three-dimensional KB primitive decays continuously to zero at its support boundary, avoiding the density jump present at m=0m=0. Its projected attenuation also has a continuous first derivative at the support boundary, keeping the corresponding projection gradients finite. Higher orders (m>1m>1) provide additional boundary smoothness while learned β\beta controls the concentration within the boundary. However, the additional smoothness does not necessarily improve the reconstruction of sharp anatomical transitions. In our experiments, higher orders yield diminishing reconstruction gains relative to their computational cost, motivating our choice of m=1m=1.

AuthorsOpenReview ↗

Response to Reviewer 922m (3/4)

Q3. Attribution of performance improvements (additional evidence). The matched representation ablation in Table 2 reports a relatively small improvement from replacing Gaussian primitives with learnable KB primitives, whereas the complete framework achieves a substantially larger improvement over R2- Gaussian. Could the authors provide a controlled component-wise ablation separating the effects of the primitive family, learnable concentration, reconstruction-guided population control, and rendering implementation? In particular, how does a Gaussian-based implementation perform when equipped with the same reconstruction-guided population control and optimization settings as KB-Splatting? This comparison would help establish whether the main improvements arise from the proposed representation or from the broader optimization pipeline.

We expanded the controlled study to 24 configurations across 12 LoopX scans at 50% and 25% views with three random seeds, yielding 1,584 evaluated runs. We update Table 2 in our paper and also provide the ablation study below for your convenience.

Furthermore, Gaussian here denotes our implemented Gaussian kernel control, not the R$^2$-Gaussian method. Our previous Gaussian control reused our KB rasterization pipeline, which we found confounded our experiments. In contrast, the current control uses a separate implementation built around Gaussian line integration, with accurate rendering, optimization, and population-control settings that yield experimentally correct quantitative results. This, along with a more expansive ablation study, explains the updated quantitative results in Table 2.

We agree that FDK-based allocation introduces a structural prior, and our expanded ablation explicitly tests its contribution. Our residual guidance proposes where to allocate primitive capacity; it does not change the analytic KB profile or introduce a reference-volume fitting loss. Nevertheless, its influence on primitive placement can affect the reconstructed anatomy, and FDK-based initialization also influences the starting representation. We therefore do not treat reconstruction guidance as inconsequential.

Table 2 separates these effects through four conditions: FDK initialization with guided allocation, the same initialization with unguided allocation, random geometric initialization with unguided allocation, and per-ray rendering. Without any reconstructed-volume prior, the learned β\beta KB pipeline exceeds the Gaussian control by 0.95 dB in mean attenuation PSNR. Thus, the advantage of this KB configuration does not depend entirely on FDK guidance. We also compare physical per-ray rendering for each primitive family (evaluated along the actual source-to-detector ray for each detector pixel) with the standard center-ray approximation that uses the viewing direction through the center of each primitive (Zwicker et al., 2002). During this comparison, we keep the projection data, losses, population budgets, and training schedules fixed.

Together, these experiments address the contributions of the primitive family, learned concentration, reconstruction guidance, and rendering approximation.

Table 2: Ablation results for Gaussian and KB primitives.

MethodPSNR ↑LPIPS ↓Time ↓Count
Full / guided
Gaussian45.440.07541.8225k
Fixed TG45.440.07537.3225k
Learned TG45.490.07437.1225k
Fixed-β KB46.810.05915.2225k
Learned-β KB (half)46.700.0589.2113k
Learned-β KB46.860.05616.3225k
Unguided allocation
Gaussian44.870.07940.9225k
Fixed TG44.890.07836.4225k
Learned TG44.920.07836.2225k
Fixed-β KB46.220.0659.9225k
Learned-β KB (half)46.160.0646.7113k
Learned-β KB46.400.06010.8225k
No reconstruction prior
Gaussian45.380.07326.9225k
Fixed TG45.410.07222.9225k
Learned TG45.460.07222.2225k
Fixed-β KB46.160.0656.9225k
Learned-β KB (half)45.910.0654.7113k
Learned-β KB46.330.0617.7225k
Per-ray rendering
Gaussian45.500.07541.8225k
Fixed TG45.500.07535.8225k
Learned TG45.530.07435.7225k
Fixed-β KB46.810.05926.2225k
Learned-β KB (half)46.720.05814.7113k
Learned-β KB46.860.05626.8225k
AuthorsOpenReview ↗

Response to Reviewer 922m (4/4)

Q4. Reconstruction accuracy and efficiency under matched evaluation conditions (clarification). The reported reference volumes are generated using analytical or proprietary reconstruction procedures rather than independent attenuation ground truth. How sensitive are the reported improvements to the evaluation reference and its reconstruction artifacts? Additionally, could the authors clarify whether the approximately twofold training speed advantage is measured under fully matched hardware, primitive budgets, and implementation conditions? Reporting final primitive counts and support-overlap statistics would help connect the claimed compact-support advantage to the observed runtime improvements.

We agree that our dataset evaluation volumes are reference reconstructions, and we now make this explicit in our figures and captions. We also emphasize the clinical relevance of our dataset. Both the Loop-X and experimental CT machine data provide measured X-ray projections, paired with their scanner-reconstructed volumes and acquisition geometry. Furthermore, our dataset contains recorded intensity conversions, allowing reconstruction methods to be evaluated in the coordinates and units of the actual image acquisitions. These features make our dataset particularly relevant to developing reconstruction methods for clinically relevant imaging scenarios.

On training speed, our expanded ablation study separates the training-time comparison between our KB primitive and the Gaussian splatting primitive from their respective complete pipelines. As stated in Table 2, across 11 LoopX cases at 50% and 25% views with three seeds, the full/guided learned-β\beta KB configuration requires an average training time of 16.4 minutes, compared with 42.3 minutes for the Gaussian control, corresponding to a 2.58×2.58\times training speedup. We attribute this non-negligible speedup to early rejection of noncontributing primitive-pixel pairs in the KB rasterizer. KB's exact support boundary allows these pairs to be discarded before calculating their projected attenuation contributions or accumulating gradients. This can reduce unnecessary arithmetic during both the forward and backward passes.

Finally, R$^2$-Gaussian belongs to the same Gaussian primitive family as our control, so KB's early support rejection is also a plausible contributor to its efficiency relative to R$^2$-Gaussian. For benchmark comparisons, both methods are trained on NVIDIA L40S GPUs with the same random seed. Both methods also use the center-ray (local affine) approximation developed in EWA splatting (Zwicker et al., 2002). Their implementations of rendering, population control, and optimization remain method-specific. As such, the approximately 2×2\times speedup measures the complete pipelines and cannot be assigned entirely to the primitive. The table below reports the actual final counts and support-overlap statistics alongside training times. In particular, we report the average number of contributing primitives per detector pixel to quantify how footprint overlap affects the rasterizer workload.

Across the four LoopX view settings, KB-Splatting has 2.3–3.5$\times$ fewer contributing primitives per detector pixel than R2-Gaussian, despite ending with more primitives overall. Its training time is 1.6–2.0$\times$ shorter. The comparison with X2-Gaussian holds the final primitive count equal at each view setting: KB-Splatting has 2.7–4.2$\times$ fewer contributors per pixel and trains 2.2–2.8$\times$ faster. These results show why final primitive count alone does not describe rasterization cost. The lower footprint overlap is consistent with less work per detector pixel.

Table 3: LoopX final primitive counts, detector-pixel contributors, and training times.

ViewsMethodFinal countContributors/pixelTime (min)
50%X-Field201,508174.466.34
50%R2-Gaussian186,0371,835.0016.39
50%X2-Gaussian225,4682,169.6122.15
50%KBS225,468814.2710.29
25%X-Field201,519175.196.35
25%R2-Gaussian183,2931,807.3916.23
25%X2-Gaussian222,2252,121.4121.85
25%KBS222,225789.6110.18
10%X-Field201,401178.956.45
10%R2-Gaussian178,4401,743.9615.96
10%X2-Gaussian217,1382,036.1521.19
10%KBS217,138666.089.10
5%X-Field201,284183.186.57
5%R2-Gaussian170,2151,601.5814.76
5%X2-Gaussian210,9711,953.5020.50
5%KBS210,971461.017.39
Reviewer_WUqkOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes to form a 3D reconstruction of a CT volume from a limited number of X-ray projections, in order to limit X-ray radiation exposure to the patient (aka cone-beam computed tomography (CBCT) reconstruction). Previous SOTA methods use NeRFs or Gaussian primitives, whilst this paper proposes to use Kaiser-Bessel (KB) functions/primitives instead.

One of the main differences between KB and Gaussians is that KBs remain exactly zero outside their boundary, whereas Gaussians technically extend infinitely. It is hypothesised that using KBs yields sharper, higher-fidelity reconstructions in regions of abrupt attenuation change, while requiring fewer Gaussians and lower optimisation cost than using a Gaussian representation.

The paper evaluates on 15 scans acquired with two cone-beam CT systems. Whilst the SSIM improvement is minimal, PSNR improves ~1dB, and the time taken to form reconstructions is still respectable at ~10 minutes (and an improvement over one of the current SOTA models which takes ~16.5 minutes)

Critical strengths
  1. To the best of my knowledge, the authors are the first to use Kaiser-Bessel functions (aka Blobs) in a gaussian splatting style framework, for CT reconstruction. Whilst this is an interesting idea worth exploring, the benefits of these new primitives do not appear to be very convincing, at least in the paper's current form.
Critical weaknesses

Primarily, this paper needs more thorough experiments to convince readers that it is worth switching from gaussian splatting to KB splatting for CBCT.

  1. The PSNR improvements in Table 1 are only ~1 dB, whilst the SSIM improvement is practically non-existent (this is arguably evident by the fact that SSIM is reported to 5 decimal places). KB-Splatting also takes 3x as long as X-Field. It is also only evaluated on 15 scans. Testing on an additional dataset could potentially show off KB-Splatting more, and help make it more convincing, as would using another visual metric (e.g. LPIPS).

  2. Whilst they do compare to other well known CT-specific Gaussian Splatting methods (as well as CT-specific NeRFs), they do not compare to current SOTA. R2 Gaussian (NeurIPS 2024) has since been surpassed by X2-Gaussian (ICCV 2025, from the same authors), and achieves improvements of ~2 dB, which would seem to surpass KB-splatting's current 1dB improvement over R2 Gaussian. So it would be very useful to the community if the authors compared their method against R2 gaussian's newer version, X2-Gaussian. Similarly, DIF-Gaussian (Y. Lin et al., Learning 3D Gaussians for Extremely Sparse-View Cone-Beam CT Reconstruction, MICCAI 2024) is able to use extremely sparse-view CT scans (≤10\leq 10 projections), and still achieves high visual quality reconstructions. A comparison to this when DIF-Gaussian is given the higher number of views the other methods in Table 1 is given would be very interesting and relevant.

  3. In Table 2, comparing the Matched Gaussian with a KB with a learned βi\beta_i (as is used in the final version), the HU PSNR and Attenuation PSNR improvement is only 0.12 dB and 0.117 dB respectively, for similar reconstruction time. Whilst it is a positive improvement, readers may question whether it is worth the slightly added complexity of using KBs, and worth the effort to transition from 3DGS which is now a well researched field, to a new primitive type (KB).

Questions for authors
  1. Are you able to include LPIPS as a visual metric, this could make the argument for using KBs more convincing?
  2. It would also help to evaluate on more than 15 scans, especially given that on these 15 scans the quantitative improvement is not that significant. Perhaps include the FIPS dataset as R2 Gaussian also test on.
  3. Are you able to provide an explanation as to why exactly KBs are better than Gaussian primitives for CBCT, because the data is not very convincing, and I could not find a good explanation within the paper?
4 comments
AuthorsOpenReview ↗

Response to Reviewer WUqk (1/4)

Thank you for finding our idea interesting and recognizing the novelty of our use of the Kaiser-Bessel kernel as a functional primitive in splatting for CT reconstruction.

Q1. The PSNR improvements in Table 1 are only $\sim$1 dB, whilst the SSIM improvement is practically non-existent (this is arguably evident by the fact that SSIM is reported to 5 decimal places). KB-Splatting also takes 3x as long as X-Field. It is also only evaluated on 15 scans. Testing on an additional dataset could potentially show off KB-Splatting more, and help make it more convincing, as would using another visual metric (e.g. LPIPS).

We agree that the small SSIM differences should not be overstated. SSIM can approach saturation near unity in high-fidelity reconstructions, making global scores less sensitive to remaining local differences. This limitation has also been documented in tomographic reconstruction Frixou et al., 2026. We retain SSIM for consistency with prior work, but interpret it alongside reconstruction errors and qualitative comparisons rather than as a sufficient measure of reconstruction quality. Following your feedback, we now report SSIM to three decimal places.

We have also expanded the quantitative reports in Tables 7–10 to report each anatomy and inanimate object in the LoopX dataset, rather than only aggregate averages. Across the 50%, 25%, 10%, and 5% view settings, KB-Splatting takes 1.44$\times$ the average training time of X-Field, while improving attenuation PSNR by 3.35 dB under the same evaluation protocol. The accompanying qualitative comparisons in Figures 11-18 illustrate the differences in anatomical detail that are not apparent from aggregate quantitative metrics alone.

Following your recommendation, we additionally evaluate on the FIPS dataset and report LPIPS in all quantitative reports. LPIPS measures perceptual differences using deep features and provides a complementary assessment to PSNR and SSIM ( Zhang et al., 2018).

AuthorsOpenReview ↗

Response to Reviewer WUqk (2/4)

Q2. Could the authors compare KB-Splatting against X$^2$-Gaussian (ICCV 2025), which reports a $\sim$2 dB improvement over R$^2$-Gaussian compared with KB-Splatting's $\sim$1 dB gain? Additionally, could they compare against DIF-Gaussian (MICCAI 2024), which achieves high visual quality with $\leq$10 projections, giving DIF-Gaussian the same higher number of views used by the other methods in Table 1?

We follow your suggestion and evaluate X$^2$-Gaussian and DIF-Gaussian on LoopX, AIAI, and FIPS using X-ray projection views matched with our KBS method. X$^2$-Gaussian is a temporally aware dynamic 4D CBCT method. As such, we adapt its static reconstruction stage to our measured source--detector geometry. Our Table 1 reports the quantitative comparisons with X$^2$-Gaussian. Averaged across the 50%, 25%, 10%, and 5% view settings, KBS improves attenuation PSNR by 0.51 dB on LoopX, 1.17 dB on AIAI, and 2.36 dB on FIPS.

DIF-Gaussian, on the other hand, learns from supervised projection–volume pairs, whereas KBS reconstructs each evaluation case directly from its measured projections, using reference volumes only for evaluation. We adapt DIF-Gaussian’s circular-trajectory mapping to our measured per-view source–detector geometry. Although the scene geometry passed, training exhibited output collapse: the network predicted small negative attenuation values instead of anatomical structure, producing empty reconstructions after nonnegative clipping. Diagnostics identify degenerate output normalization and background-dominated training samples as potential contributors. We test balanced sampling and normalization layers. Yet these changes recover coarse anatomy but leave excessive background signal and insufficiently render fine details. Consequently, we do not obtain reliable quantitative and qualitative results. This unresolved instability, together with its dependence on supervised volume targets, made our DIF-Gaussian adaptation unsuitable for the present projection-supervised comparison.

AuthorsOpenReview ↗

Response to Reviewer WUqk (3/4)

Q3. In Table 2, comparing the Matched Gaussian with a KB with a learned βi\beta_i (as is used in the final version), the HU PSNR and Attenuation PSNR improvement is only 0.12 dB and 0.117 dB respectively, for similar reconstruction time. Whilst it is a positive improvement, readers may question whether it is worth the slightly added complexity of using KBs, and worth the effort to transition from 3DGS which is now a well researched field, to a new primitive type (KB).

We agree that the original comparison provides limited evidence for the benefit of the primitive itself. We therefore expand the ablation to 12 LoopX cases with three random seeds at both 50% and 25% views. We compare six variants—Gaussian, fixed and learned Truncated Gaussian (TG), fixed-β\beta KB, and joint-update learned-β\beta KB at full and half primitive budgets—under four configurations: full/guided, unguided allocation, no reconstruction prior, and per-ray rendering. This comprises 24 method–configuration combinations and 1,584 runs, with 792 runs per view percentage.

Table 2 compares Gaussian, fixed and learned truncated Gaussian (TG), fixed-β\beta KB, and joint-update learned-β\beta KB across 12 LoopX cases with three seeds at 25% views. Additional configurations test unguided allocation, random initialization without a reconstruction prior, and per-ray integration. We construct the Gaussian control from R$^2$-Gaussian, retaining its renderer, voxelizer, and optimization objective while adding FDK-residual-guided birth proposals and matching training views, seeds, iteration budgets, and primitive caps. These experiments assess whether KB's advantage persists across changes in reconstruction guidance and rendering.

Under the full/guided configuration, learned-β\beta KB improves mean attenuation PSNR over Gaussian by 1.42 dB and reduces LPIPS from 0.075 to 0.056. Introducing fixed or learned Gaussian truncation improves Gaussian PSNR by at most 0.05 dB. Fixed-β\beta KB achieves 46.81 dB, compared with 46.86 dB for learned-β\beta KB. Figure 5 complements these results with matched hand and phantom-neck regions, where KB variants preserve clearer bone boundaries and internal separations. Halving each case's KB primitive cap reduces the average population from 225k to 113k with a PSNR decrease of only 0.16 dB; this smaller model still exceeds the full-budget Gaussian control by 1.26 dB.

The component comparisons also clarify the accuracy–runtime trade-off. Removing allocation guidance lowers learned-β\beta KB's PSNR by 0.46 dB, and removing FDK initialization further lowers it by 0.07 dB. Per-ray integration leaves PSNR and LPIPS unchanged, but increases recorded training time from 16.3 to 26.8 minutes; center-ray rendering ( Zwicker et al., 2002) therefore reduces training time by approximately 39% in this comparison. Gaussian primitive family and KB retain the same optimization settings. As such, these results attribute the metric improvements to the KB primitive.

AuthorsOpenReview ↗

Response to Reviewer WUqk (4/4)

Q4. Are you able to provide an explanation as to why exactly KBs are better than Gaussian primitives for CBCT, because the data is not very convincing, and I could not find a good explanation within the paper?

Our motivation for KB primitives is that they separate spatial support from profile concentration: the support parameters determine where a primitive contributes, while β\beta controls how attenuation is distributed within that support. In a Gaussian, changing its scale changes both its spatial extent and decay. KB therefore offers additional flexibility for representing localized attenuation transitions without extending contributions indefinitely into neighboring regions. However, compact support alone does not establish superiority, so we expanded Table 2 to include Gaussian, fixed and learned truncated Gaussian, and fixed and learned-β\beta KB across 12 LoopX cases. At 25% views, guided learned-β\beta KB improves mean PSNR over Gaussian by 1.42 dB and reduces LPIPS from 0.075 to 0.056, whereas Gaussian truncation improves PSNR by at most 0.05 dB. Fixed-β\beta KB already achieves 46.81 dB, close to the 46.86 dB obtained with learned-β\beta KB and joint update. These findings suggest that the observed advantage involves more than simply truncating Gaussian support or learning an additional concentration parameter.

The additional controls help distinguish this representation advantage from the contribution of reconstruction guidance and computational budget. Removing allocation guidance lowers learned-β\beta KB's PSNR by 0.46 dB, while additionally removing FDK initialization lowers it by another 0.07 dB, showing that guidance contributes to the final result. Nevertheless, reducing KB's average population from 225k to 113k costs only 0.16 dB, and this smaller model still exceeds the full-budget Gaussian control by 1.26 dB. Figure 5 illustrates the corresponding preservation of bone boundaries and internal separations in matched hand and phantom-neck regions. Together, these results support KB as an effective and compact representation for CBCT reconstructions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Any Last-minute Feedback and Thanks to All the Reviewers

Dear reviewers,

As the discussion period draws to a close, we welcome any final feedback. Thank you for your thoughtful suggestions. We have worked to address your concerns in the attached revised manuscript, where new content is marked in blue for ease of review. We now submit the appendix separately from the main paper due to size restrictions.

The revision adds references and figures and refines the text to present the work more accurately, clearly, and concisely within the page limit. We appreciate the time and care you have given our paper.

Wishing you a happy Thanksgiving.

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.