acceptodds
Under review as a conference paper at ICLR 2027

Beyond Instant-NGP: A Unified Variable-Order B-spline Framework for Sparse-View CT Reconstruction

Abstract

Achieving high-quality CT reconstruction from sparse measurements is essential for reducing radiation exposure without compromising diagnostic image quality. However, the limited number of projections makes image reconstruction severely underdetermined. In this context, Instant-NGP has emerged as an efficient implicit neural representation for sparse-view CT reconstruction, yet its representational properties underlying its effectiveness remain poorly understood. In this work, we provide a systematic mathematical analysis of the original methodology and specific implementation of Instant-NGP, and show that its multi-resolution grid encoding is equivalent to a multi-resolution linear B-spline representation, in which the CT image is approximated by multivariate polynomials constructed from locally supported linear B-spline basis functions. We further show that the approximation capability of this linear B-spline representation is fundamentally limited for piecewise-smooth signals, potentially leading to reconstruction errors and artifacts. Motivated by this observation, we generalize the linear representation to a unified variable-order B-spline framework, allowing higher-order basis functions to provide more accurate local approximation while retaining the efficiency of hash encoding. We theoretically establish that, under appropriate smoothness conditions, increasing the spline order tightens the bound on the optimal approximation error. Extensive quantitative and qualitative experiments on sparse-view CT reconstruction validate our analysis, showing that higher-order B-splines consistently improve reconstruction quality over the linear formulation, with quartic B-splines achieving the strongest overall performance. The code will be open-sourced upon acceptance of the paper.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
CT Reconstruction, Variable-Order, B-spline, Sparse-View

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7XdROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper establishes a connection between Instant-NGP and B-spline functions and proposes a variable-order B-spline encoding method for sparse-view CT reconstruction. The authors provide theoretical analysis of approximation error and evaluate the proposed method on 2D CT reconstruction tasks.

Critical strengths
  1. Theoretical analysis of B-spline encoding. The paper provides a useful mathematical analysis of the relationship between Instant-NGP and linear B-splines. It further shows that, under suitable smoothness conditions, increasing the B-spline order can lead to a tighter upper bound on the optimal approximation error. This provides a theoretical motivation for using higher-order B-splines.

  2. Consistent improvements across different numbers of views. The experiments show that higher-order B-spline encoders consistently improve reconstruction quality across different view settings. For example, the quartic B-spline achieves higher PSNR than the linear B-spline in the reported lung CT experiments. The CUDA implementation also helps maintain computational efficiency.

Critical weaknesses
  1. Limited comparison with recent methods. The experiments mainly compare the proposed method with traditional reconstruction methods and earlier neural representation methods, such as SIREN, RDFN, and Instant-NGP. Although most recent works on improving neural representations for CT reconstruction focus on 3D CT, NAB (ICLR 2026) studies an alternative coordinate encoding method for sparse-view 2D CT reconstruction. However, the paper does not compare with NAB or clearly discuss the differences between the two methods.

  2. The performance improvements appear relatively modest. Although higher-order B-spline encoders consistently improve reconstruction quality, the gains over the linear B-spline encoder are relatively modest. On the lung dataset, the quartic B-spline improves PSNR by approximately 0.38–0.52 dB across different view settings.

Questions for authors
  1. Can the proposed B-spline encoder be integrated into other neural CT reconstruction methods? It would be interesting to explore whether the proposed B-spline encoder can serve as a plug-in replacement for the original coordinate encoder in other frameworks. Consistent improvements across different methods, even if the resulting methods do not achieve the best overall performance, would help demonstrate the compatibility and general applicability of the proposed encoder.

  2. Could the authors clarify the relationship between the proposed method and other encoding methods? In particular, how does the variable-order B-spline encoding differ from Freq-NAF and NAB? A clearer discussion would help readers better understand the novelty and advantages of the proposed method. If feasible, a comparison with NAB under the same experimental setting would also help better evaluate its practical value.

1 comment

I appreciate the authors' efforts and the additional experiments on 95 independent patients. Comparing the proposed method with recent SOTA methods (e.g., NAB) would be a direct way to show its value. If such a comparison is not available, the authors should at least show its value by demonstrating that it works well across different neural CT reconstruction frameworks. However, the new experiments only show that the method generalizes to different patients, not to different reconstruction frameworks. Therefore, I believe the value of the proposed method has not been fully demonstrated, and I maintain my original rating.

Reviewer_v7G3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates Instant-NGP from the perspective of B-spline approximation theory and proposes a variable-order B-spline framework for sparse-view CT reconstruction. The authors first establish the equivalence between Instant-NGP's multilinear grid interpolation and tensor-product linear B-spline representations. They further analyze the effects of hash collisions and coordinate offsets, and provide an approximate polynomial interpretation of the network output.

Motivated by the limited approximation capability of linear B-splines, the authors extend the encoding to higher-order B-splines and derive an order-dependent approximation error bound under smoothness assumptions. They implement five variants, ranging from constant to quartic B-splines, using CUDA-accelerated hash encoding.

Experiments on lung and liver CT datasets with simulated sparse-view projections demonstrate generally consistent improvements in PSNR and SSIM as the spline order increases. The quartic variant achieves the best overall results, with approximately 5% additional training time compared with the linear variant.

Critical strengths

1. A systematic B-spline interpretation of Instant-NGP.

The paper provides a coherent mathematical interpretation of Instant-NGP's encoding mechanism using locally supported B-spline basis functions. Beyond the basic interpolation equivalence, the authors examine implementation-specific details, including hash collisions and coordinate offsets, and discuss their implications from an approximation-theoretic perspective.

This analysis is valuable because it provides an interpretable framework for understanding multi-resolution hash encoding and offers a systematic starting point for modifying its representation.

2. Consistent empirical improvements with limited computational overhead.

The paper systematically evaluates five B-spline variants across different projection counts and hashing configurations. The results generally show improved reconstruction quality with increasing spline order, supported by both quantitative metrics and qualitative comparisons.

For example, Table 1 reports PSNR improvements of approximately 0.38-0.52 dB for quartic B-splines over hashed linear B-splines on the lung dataset. Importantly, Table 4 indicates that these improvements incur only about 5% additional training time.

The relatively consistent performance gains and low computational overhead make the proposed extension practically interesting, even though the precise mechanism behind the improvements remains insufficiently established.

Critical weaknesses

1. Insufficient motivation for higher-order B-spline encoding and limited evidence for the claimed improvement mechanism.

The central motivation of the paper is that the limited approximation capability of linear B-splines motivates replacing Instant-NGP's interpolation with higher-order B-splines. However, it is not sufficiently established that this limitation is a primary source of reconstruction errors in the full Instant-NGP architecture.

In particular, the approximation limitation illustrated in Appendix A.7 is demonstrated in an encoder-only setting, whereas Instant-NGP combines linear interpolation with a nonlinear neural network. As discussed in Section 3.6, the network can already construct higher-degree interactions from linear B-spline features. Therefore, the analysis does not directly establish why increasing the spline order at the encoding stage is needed, or why it should be preferred over improving the expressiveness or regularization of the existing architecture.

Furthermore, increasing the spline order simultaneously changes the spatial support and smoothness of the interpolation basis. Consequently, the observed improvements may arise from stronger spatial smoothing or implicit regularization rather than the claimed advantage in approximation order. The current experiments do not disentangle these factors or compare against simpler alternatives, such as regularized linear hash encoding or other locally supported interpolation kernels.

The approximation error bound in Eq. (16) also provides only a limited justification for the proposed design. It characterizes the optimal approximation error under smoothness and representability assumptions, rather than the reconstruction error arising from sparse projection measurements. Moreover, the order-dependent constants and the saturation of approximation rates at the target smoothness limit make it difficult to infer a strictly monotonic improvement with spline order at a fixed resolution.

This is a central weakness because the paper's main methodological contribution is motivated by the purported approximation limitations of linear B-splines. While the empirical results demonstrate that higher-order variants can be beneficial, the paper does not convincingly establish that the proposed modification addresses the identified limitation or that its advantages are specific to higher-order B-spline approximation.

2. Limited experimental validation and insufficient comparison with competitive alternatives.

Although the paper reports consistent improvements across different projection settings, the experimental evaluation is based on a relatively small number of independent anatomical instances. Specifically, the lung and liver datasets contain 100 and 40 slices, respectively, but these are extracted from only four lung volumes and two liver volumes.

Since slices from the same volume are highly correlated, the per-slice statistical tests do not necessarily provide strong evidence of generalization across independent patients. The evaluation is also restricted to simulated parallel-beam projections, without clearly establishing robustness to realistic measurement noise or forward-model mismatch.

Additionally, while the paper compares against several classical reconstruction and generic INR methods, it does not include sufficiently strong comparisons with recent CT-specific INR approaches, despite discussing several such methods in the related work. This makes it difficult to assess the practical significance of the reported improvements relative to the current state of the art.

Another concern is that Appendix A.10 states that the best PSNR values are used for reporting. It is unclear whether this involves selecting checkpoints based on ground-truth reconstruction quality, which would not be available in practical self-supervised CT reconstruction.

These limitations are particularly relevant because the improvements over linear B-splines, while consistent, are relatively modest. Without stronger baselines, more independent test cases, and a clearly specified evaluation protocol, it remains uncertain whether the reported gains reflect a robust and practically significant advantage.

Questions for authors

1. Motivation and necessity of higher-order B-spline encoding (clarification and additional evidence).

The paper motivates higher-order B-splines through the approximation limitations of linear B-splines. However, Appendix A.7 considers an encoder-only setting, whereas the full Instant-NGP architecture contains a nonlinear MLP.

What evidence indicates that the approximation capability of linear B-spline encoding is a primary bottleneck in the full architecture? Could comparable improvements be obtained by increasing network capacity, adjusting grid resolutions, or introducing regularization while retaining linear interpolation?

Clarifying this point would help establish whether the proposed modification addresses a fundamental representational limitation or primarily provides a more favorable inductive bias.

2. Source of the observed improvements (clarification and additional evidence).

Increasing the B-spline order changes not only the approximation order but also the spatial support and smoothness of the interpolation kernel.

Can the authors distinguish the benefits of higher-order approximation from those of increased spatial smoothing or implicit regularization? For example, comparisons against linear hash encoding with explicit smoothness regularization or alternative locally supported interpolation kernels would help isolate the mechanism.

Relatedly, does the increased smoothness affect the preservation of sharp anatomical boundaries, particularly under severe view undersampling?

3. Interpretation of the theoretical guarantee (clarification).

In Eq. (16), how do the approximation constants depend on the spline order? Does the bound guarantee an improvement at a fixed grid resolution, or does it characterize only the asymptotic approximation rate under sufficient smoothness?

Furthermore, how does this analysis relate to the actual CT reconstruction problem, where the reconstruction error also depends on projection undersampling, hash collisions, and optimization?

4. Evaluation protocol and generalization (clarification and additional evidence).

Appendix A.10 states that the best PSNR is used for reporting. Are checkpoints selected using ground-truth images? If so, how would the results change under a fixed training budget or a stopping criterion that does not require ground truth?

Additionally, could the authors report per-volume results or an analysis accounting for correlations between slices from the same volume? Evidence under realistic projection noise and comparisons with recent CT-specific INR methods would also help determine whether the reported improvements generalize beyond the current experimental setting.

Post-review comments

I appreciate the authors' detailed response and the additional experiments.

The new evaluation on 95 independent patient volumes addresses my concern about within-volume correlations. The use of held-out projection measurements for checkpoint selection also clarifies the evaluation protocol for the additional experiments. Furthermore, the comparisons across network depths, grid resolutions, and regularization strengths provide stronger empirical evidence that higher-order B-splines can improve reconstruction quality under different configurations. I acknowledge these improvements.

However, my main concern regarding the methodological motivation remains unresolved.

The authors argue that the approximation limitations of linear B-splines should persist after introducing a nonlinear MLP. This argument is based on intuition rather than an analysis of the complete representation. The additional experiments demonstrate that quartic B-splines outperform linear B-splines under the tested configurations, but they do not establish that the limited polynomial approximation order of linear B-splines is the underlying bottleneck.

Increasing the spline order simultaneously increases the spatial support and smoothness of the basis functions. The TV and Tikhonov experiments are useful, but they do not isolate these effects. Consequently, the observed improvements cannot be attributed specifically to higher-order approximation. The authors' explanation that larger support is inherent to higher-order B-splines does not resolve this concern.

The theoretical clarification also confirms the limited scope of Eq. (16). The bound concerns the best-approximation error under smoothness assumptions. It does not guarantee improved reconstruction at a fixed grid resolution and does not account for sparse-view measurements, hash collisions, or optimization. Therefore, the bound does not establish a theoretical explanation for the observed reconstruction gains.

I also find the claim of improvement without hyperparameter tuning insufficiently justified. The spline order is itself a model-design choice. Moreover, the authors use held-out projection measurements for checkpoint selection, so the assertion that validation-based parameter selection is infeasible in sparse-view CT requires further qualification. The reported regularization sweeps also contain configurations without statistically significant improvements.

Overall, the rebuttal substantially strengthens the empirical evidence and resolves my concern about statistical dependence between slices. I no longer consider that issue a major weakness.

However, my central methodological concern remains: although the additional experiments establish a consistent empirical advantage for higher-order B-splines, they do not demonstrate that the encoder-only approximation limitation remains a bottleneck in the full INR model. Nor do they disentangle higher-order approximation from the accompanying changes in spatial support and smoothness.

I maintain my rating of 2.

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

The paper evaluates the effect of using linear B-splines to interpolate the multi-resolution grids used in Instant-NGP on the sparse-view CT reconstruction problem. The paper interpolates the multi-resolution grids using splines and retrieves the spline coefficients using a hash function given the coordinates. Using a Chebyshev polynomial approximation of the nonlinearities in the MLP, they show that Instant-NGP with spline interpolation is a polynomial function of the B-spline basis functions. They further experiment with higher-order splines, give an error bound which reduces with the order of the splines, and experimentally show that the runtime overhead increases by a small amount. They also claim that hash collisions act as a regulariser during training. The splines are evaluated on lung and liver CT datasets, and the results show that the quartic B-spline gives better reconstruction in terms of SSIM and PSNR.

Critical strengths

Evaluating the efficacy of higher-order splines to interpolate feature grids.

  1. Instant-NGP avoids higher-order splines owing to their computational complexity (Müller et al.). The paper shows that, with their implementation, the computation time increases by 7 percent compared to constant interpolation, which provides the affirmative case for using higher-order splines.
Critical weaknesses

The ideas proposed in the paper have minimal evaluation.

  1. The training time of the baseline Instant-NGP is missing, as well as comparison with standard deep-learning based methods used for sparse-view CT (e.g., CoIL, Sun et al., IEEE TCI 2021) and methods with standard regularisers (e.g., TV). These would give an estimate of the state of the art in terms of accuracy and time.
  2. Hash collision as regularisation seems like an interesting idea. However, the paper doesn’t mention what kind of artefacts it introduces as collisions increase. Secondly, as the hash function only depends on the coordinates, it might introduce artefacts only at certain locations irrespective of the data, so the locations of collisions and their effect on local quality is not tested.

The connection between the theoretical statements and how they help in understanding the model is missing.

  1. The paper mentions that the multi-resolution grids are connected to wavelet spaces; however, it does not go into any detail on how these wavelet spaces can be used in the model, either for analysis or improvement.
  2. The paper mentions about the offsets used in Instant-NGP, where they were used to solve the zero-gradient issue (Müller et al.), but doesn’t go into detail on why they are important in higher-order spline interpolation or whether they change reconstruction quality.
  3. Theoretical bounds on the error might not be useful for evaluating spline interpolation, as the models are trained along with the interpolation. The network might therefore have a worse Lipschitz constant, leading to higher bounds even though higher-order interpolation is used.
Questions for authors

Clarification

  1. What is the training speed of the baseline Instant-NGP compared to the spline-based methods?

Suggested experiments:

  1. Can you evaluate the reconstruction quality against standard deep-learning methods, such as CoIL, or a U-Net trained on the dataset used in your experiments, as well as reconstruction using simple regularisers?
  2. Using oracle estimates to pick the best PSNR reconstruction can overestimate the results, as the oracle is only a proxy. One possible approach would be to tune hyperparameters, such as the number of epochs, on one dataset and report the reconstruction quality on another, or to use a train/test split.
  3. The hash-collision regularisation statement needs more experiments. How does hashing affect the reconstruction? Are there visual artefacts apart from the reported PSNR, and at what locations do the collisions occur?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.