acceptodds
Under review as a conference paper at ICLR 2027

Neural Fiber Operators: Learning on Antipodally Symmetric Spherical Fiber Functions

Abstract

Diffusion MRI (dMRI) is a critical non-invasive technique for characterizing white matter integrity. Predicting high angular resolution fiber orientation distribution (FOD) is crucial in clinical practice. Existing learning-based methods enable such FOD angular super resolution task but ignore the spherical-native geometry of FOD. To address this, we first reformulate FOD angular super resolution as a spherical operator learning problem, rather than conventional FOD regression. Based on this, we propose Neural Fiber Operator (NFO), a novel antipodal-symmetry-based neural operator, which efficiently learns the solution operator from FOD while preserving the antipodal symmetry of FOD. To exploit NFO, we design NFONet, a patch-based neural operator model for whole-brain or partial-brain FOD modeling. Extensive experiments are conducted on three complementary diffusion MRI datasets covering in-domain reconstruction, zero-shot generalization to independent cohorts and acquisition protocols, and ex-vivo biological validation. NFO consistently achieves state-of-the-art performance while using substantially fewer parameters. More importantly, NFO demonstrates robust generalization across progressively increasing distribution shifts, suggesting that the learned operator captures intrinsic properties of FOD angular super resolution. These results establish NFO as a principled framework for dMRI reconstruction and FOD modeling, and suggest that learning physically consistent solution operators provides a natural formulation for brain structure.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neural fiber, Neural operator, Computational physics, Neuroscience

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_Ln6hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Neural Fiber Operators (NFO) for angular super-resolution reconstruction of Fiber Orientation Distributions (FODs) in diffusion MRI. The authors formulate the mapping from low-angular-resolution FODs to high-angular-resolution FODs as a neural operator learning problem on the sphere. Specifically, starting from the Constrained Spherical Deconvolution (CSD) model, the authors introduce spherical differential operators and Green's functions, and derive a spectral operator through Spherical Harmonic (SH) expansion. The resulting operator acts exclusively on even-degree SH coefficients to preserve the antipodal symmetry of FODs. The proposed operator is then integrated into a patch-based architecture, termed NFONet, which predicts the high-quality FOD at the center voxel using a local 9 × 9 × 9 voxel neighborhood. The experiments include reconstruction on the HCP-YA dataset, zero-shot generalization to the MGH dataset, and cross-domain evaluation on corpus callosum data. The authors also provide tractography and structural connectome evaluations. The reported results demonstrate that the proposed method achieves competitive reconstruction accuracy against the selected baselines while requiring fewer trainable parameters.

Critical strengths

1.The authors evaluate the proposed method not only on the HCP-YA dataset but also on the MGH and corpus callosum datasets, covering different acquisition protocols. NFO consistently outperforms the selected baselines in terms of the reported angular correlation coefficient metric.

2.NFO contains only 2.74M trainable parameters, substantially fewer than FOD-Net and FOD-Swin-Net. Furthermore, the authors provide tractography and structural connectome analyses, extending the evaluation beyond voxel-level SH coefficient prediction.

Critical weaknesses
  1. The core operator introduced in Eq. (20) is mathematically very similar to the GSNO operator presented in Eq. (12) of Tang et al., Generalized Spherical Neural Operators: Green's Function Formulation (ICLR 2026). Moreover, the emphasized antipodal symmetry and even-degree SH representation are established properties of FOD modeling, while patch-based learning has already been adopted in FOD-Net.

2.The experimental evaluation primarily compares NFO against FOD-Net (2022) and FOD-Swin-Net (2024). Although these are reasonable historical baselines, comparisons limited to conventional CNN- and Transformer-based approaches are insufficient for a paper whose central contribution is a spherical neural operator. In particular, the ICLR 2026 work on GSNO has already evaluated FOD-SFNO, ESCNN, and FOD-GSNO on the same FOD angular super-resolution task. These methods are highly relevant to the present work in terms of both the target problem and operator design, yet they are omitted from the experimental comparisons. Consequently, the current results only demonstrate that NFO outperforms the selected conventional baselines.

3.The authors emphasize the efficiency of the proposed model. However, on the HCP-YA dataset, NFO requires 282.86 s/sample for inference, compared with 161.36 s/sample for FOD-Net and 7.14 s/sample for FOD-Swin-Net. Despite having fewer trainable parameters, NFO exhibits substantially slower inference. Therefore, the paper should clearly distinguish parameter efficiency from computational efficiency and avoid conflating the two.

Questions for authors
  1. The proposed operator in Eq. (20) is mathematically very similar to the spherical spectral operator presented in Section 4.1, Eq. (12), of Tang et al., Generalized Spherical Neural Operators: Green's Function Formulation (ICLR 2026). Beyond restricting the computation to even-degree SH coefficients, what are the substantive innovations of NFO at the operator level compared with existing SFNO/GSNO formulations?

2.Why are FOD-SFNO, FOD-GSNO, and ESCNN excluded from the experimental comparisons? Given that these methods have already been evaluated on the same FOD angular super-resolution task, could the authors provide direct comparisons under consistent data splits, preprocessing procedures, and training configurations?

3.How would the model perform if the NFO block were replaced with an existing spherical spectral operator within the same NFONet architecture, while retaining the identical even-degree SH representation? To what extent do the reported performance improvements arise from the proposed operator design, rather than from the overall network architecture and training configuration?

3 comments

We sincerely thank the reviewer for the constructive feedback and for recognizing the reconstruction accuracy, cross-domain generalization, parameter efficiency, and downstream tractography and connectome evaluations of NFO.


1. Theoretical and Architectural Novelty (Weakness 1 & Question 1)

We thank the reviewer for this important question. We would like to clarify that the main contributions of NFO lie in its fiber operator formulation and dedicated nonlinear operator architecture, rather than the degree-wise spectral multiplication alone.

Theoretically, SFNO and GSNO establish general spherical operator formulations, whereas NFO specifically addresses the CSD-motivated antipodally symmetric fiber system. We formulate FOD angular super-resolution as a spherical function-to-function mapping and derive its Green's-function-based representation within the even-degree SH subspace (Section 4.2). The learned kernel is explicitly parameterized on this subspace to preserve the antipodal symmetry of FODs. We agree that antipodal symmetry and even-degree SH representations are established properties of FODs. Our contribution is not the discovery of these properties, but their incorporation into a fiber-specific operator formulation. The spectral computation follows the established SO(3)-equivariant convolution assumption in Eq. (14), leading to the degree-wise parameterization in Eq. (20).

Architecturally, the spectral multiplication represents only one component of an NFO block. Each block further integrates spherical spectral-spatial transformations, 1×1 channel interactions, and nonlinear activation. Multiple NFO blocks are composed into NFONet, progressively aggregating local spherical fiber representations to reconstruct the target FOD. In contrast, FOD-SFNO and FOD-GSNO in Tang et al. (2026) adapt FOD-Net by replacing its convolutional operations with the corresponding spherical operators (spectral multiplication without spatial transformations&operations). These baselines evaluate spherical operator parameterizations within an existing reconstruction architecture, whereas NFO develops dedicated nonlinear operator blocks and a complete multi-block operator network.

We have clarified these theoretical and architectural distinctions in the revised manuscript (Related Work, Methodology and Appendix B.1.4), with the corresponding changes highlighted in red.

2. Additional Spherical Operator Comparisons (Weakness 2 & Question 2)

We appreciate the suggestion to include more directly relevant spherical operator baselines. The original evaluation focused on established FOD reconstruction architectures, FOD-Net and FOD-Swin-Net. To strengthen the comparisons, we have added SDNet (2024), FOD-SFNO, and FOD-GSNO (2026).

In particular, FOD-SFNO and FOD-GSNO were previously evaluated on the same FOD angular super-resolution task. We implemented these methods following the FOD-Net-based adaptations described in Tang et al. (2026), using consistent data splits, preprocessing, training, and evaluation protocols.

The following table summarizes the mean ACC on held-out HCP-YA and zero-shot MGH/CC datasets:

MethodHCP-YA ACC ↑MGH ACC ↑CC ACC ↑
FOD-Net0.68330.64170.7894
FOD-Swin-Net0.66520.42980.4089
SDNet0.73670.67630.8243
FOD-SFNO0.70690.67030.8165
FOD-GSNO0.71670.67550.8193
NFO0.75330.68270.8333

NFO achieves the highest mean ACC across all three datasets, outperforming FOD-SFNO and FOD-GSNO in both held-out reconstruction and zero-shot generalization. Compared with the strongest competing method, SDNet, NFO improves mean ACC by 0.017, 0.006, and 0.009 on HCP-YA, MGH, and CC, respectively. The revised tables and Appendix provide complete tissue-wise ACC/RMSE results and additional statistics. Regarding ESCNN, the original method of Snoussi and Karimi (2025) estimates FODs directly from reduced dMRI measurements, whereas our task learns the mapping from LAR-FOD to HAR-FOD. In this revision, we prioritize the directly relevant FOD-SFNO and FOD-GSNO baselines and have clarified the relationship to ESCNN in Related Work.

These additional comparisons and their implementation details have been incorporated into the revised manuscript (Section 5.1), with the changes highlighted in red.

3. Operator Versus Architecture Contributions (Question 3)

We thank the reviewer for this insightful suggestion. We agree that distinguishing the contribution of spectral parameterization from that of the complete NFONet architecture is important.

We would like to clarify that an NFO block is not equivalent to its degree-wise spectral multiplication. It also incorporates spherical transformations, nonlinear activation, and channel interactions. Replacing the entire block with a bare spectral multiplier would simultaneously remove these components, making it difficult to isolate the contribution of spectral parameterization. Moreover, the equivariant spectral multiplier in Eq. (20) follows the same degree-wise parameterization principle as SFNO on the even-degree SH subspace. Replacing this multiplier with its corresponding SFNO form would therefore not introduce a distinct spectral operation.

To directly investigate an alternative operator parameterization within the same architecture, we implemented NFO + GSNO Correction, incorporating the generalized correction from GSNO into the spectral branch while retaining the remaining NFONet architecture and training protocol. We also conducted component-wise ablations to evaluate the contributions of the spectral branch, 1×1 convolution branch, and residual branch:

ModelHCP-YA ACC ↑MGH ACC ↑CC ACC ↑
NFO w/o Spectral Branch0.68910.63180.4956
NFO w/o 1×1 Convolution Branch0.75070.68230.8328
NFO w/o Residual Branch0.74940.67880.8296
NFO0.75330.68270.8333
NFO + GSNO Correction0.75390.68340.8325

Removing the spectral branch decreases ACC by 0.0642, 0.0509, and 0.3377 on HCP-YA, MGH, and CC, respectively. This demonstrates its substantial contribution, particularly under the challenging in-vivo to ex-vivo domain shift. Removing the residual branch also consistently reduces reconstruction accuracy. Introducing the GSNO generalized correction yields marginal changes in ACC and does not provide a consistent improvement over the original spectral parameterization in these experiments.

Together with the FOD-SFNO/FOD-GSNO comparisons, these controlled experiments provide complementary evidence regarding the contributions of spectral learning and the complete nonlinear NFO architecture.

We have included the new results and corresponding analysis in the revised manuscript (Section 5.5), with all additions highlighted in red.

4. Parameter Efficiency and Inference Time (Weakness 3)

We appreciate the reviewer's observation and agree that parameter efficiency and computational efficiency should be clearly distinguished.

Although NFO uses fewer parameters than FOD-Net and FOD-Swin-Net, its repeated spherical spectral transformations and nonlinear operator blocks introduce additional computational overhead. Furthermore, NFO performs voxel-centered reconstruction using local neighborhoods, whereas FOD-Swin-Net employs a more efficient patch-wise reconstruction strategy.

NFO uses only 2.74M parameters, compared with 19.59M for FOD-Net, 48.17M for FOD-Swin-Net, and 74.93M for SDNet. Compared with SDNet, NFO uses approximately 27× fewer parameters and achieves approximately 10× faster inference on HCP-YA, while obtaining best performance across all three datasets.

We have therefore revised the efficiency-related claims to emphasize the accuracy--efficiency trade-off, rather than suggesting universal superiority in inference speed. The computational limitations and runtime comparisons are now explicitly discussed in the revised manuscript (Abstract, Introduction, and Experiments).


We sincerely thank the reviewer again for the valuable and constructive suggestions. We have clarified the theoretical formulation and architectural contributions, expanded the comparisons with existing spherical operators, conducted controlled operator-level and component-wise ablations, and strengthened the discussion of computational efficiency. All corresponding changes are highlighted in red in the revised manuscript.

Dear Reviewer Ln6h,

We hope this message finds you well. We have added experiments and clarifications in response to your constructive feedback, especially the theoretical and architectural novelty, spherical operator comparisons and controlled ablation experiments.

As the discussion phase is nearing its end, we would greatly appreciate any further feedback you may have. Thank you again for your time and consideration.

Best regards,

The Authors

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

This paper studies the recovery of high-quality fiber orientation distributions (FODs) from low-angular-resolution FODs reconstructed using a reduced number of diffusion measurements. The authors formulate the task as a function-to-function mapping on the sphere and propose the Neural Fiber Operator (NFO). The model incorporates information from a central voxel and its three-dimensional neighborhood, learns degree-wise spectral weights in the spherical harmonic domain, and combines these with channel interactions to predict the central voxel’s FOD. Antipodal symmetry is incorporated as a key structural prior and preserved through an even-degree spherical harmonic representation. The evaluation includes in-distribution reconstruction, zero-shot transfer across independent cohorts and acquisition protocols, and transfer from in vivo to ex vivo data, together with downstream analyses of tractography and structural connectivity.

Critical strengths
  • Relevant and practically motivated problem. Improving FOD estimation from limited single-shell acquisitions addresses an important acquisition-time constraint.
  • The cross-domain evaluation has practical relevance. The evaluation covers in-distribution reconstruction, independent cohorts, different acquisition protocols, and transfer from in vivo to ex vivo data, providing useful evidence regarding applicability under distribution shifts.
Critical weaknesses
  • Novelty relative to existing spherical operators is insufficiently established. The derivation of the spherical spectral mapping from Green’s functions in Section 4.2 is closely related to the existing derivation in Section 4.1 of Tang et al. (2026) [1]. Furthermore, [2] explicitly describes learnable degree-wise weights in the spherical harmonic domain, weight sharing across orders within each degree, and restriction to even degrees.
  • The current experiments do not sufficiently isolate the contribution of the proposed operator design. The evaluations across acquisition protocols, independent cohorts, and in vivo and ex vivo data are valuable. However, the current results primarily characterize the performance of the complete network and provide limited insight into the roles of individual key design choices.
  • The inverse-operator interpretation relies on unverified assumptions. Equations (8)–(9) introduce an unspecified degradation operator and its inverse, although bandwidth loss can be non-invertible and noise need not admit the assumed deterministic linear model.
  • Unestablished clinical deployment benefits. The current implementation has relatively high inference time, and the study does not evaluate prospective clinical acquisitions, diagnostic utility, or end-to-end workflow reliability.
  • The "efficiency" claim is directly contradicted by the paper's own tables. The abstract states the operator "efficiently learns" and the intro motivates neural operators partly because existing approaches are "computationally inefficient for large-scale whole-brain fiber modeling." Yet in Tables 1/2/8 NFO is the slowest method

[1] Tang H, Chen H, Li C. Generalized Spherical Neural Operators: Green’s Function Formulation. ICLR, 2026. [2] Snoussi H, Karimi D. Equivariant spherical CNNs for accurate fiber orientation distribution estimation in neonatal diffusion MRI with reduced acquisition time. Frontiers in Neuroscience, 2025, 19:1604545. DOI: 10.3389/fnins.2025.1604545.

Questions for authors
  • Which assumptions ensure that DxD_x is linear, invertible, and compatible with the proposed spectral form? Should the learned mapping instead be interpreted as a regularized or statistical inverse when degradation includes noise and information loss?
  • Do existing supplementary results isolate the contributions of the operator and architecture? In particular, are there matched comparisons with even-degree SFNO, SH regression, or variants without residual and neighborhood processing?
  • Why is NFO ~40–100× slower than FOD-Swin-Net?
5 comments

We sincerely thank the reviewer for the detailed and constructive assessment. We appreciate the recognition of the practical relevance of FOD reconstruction and the extensive cross-domain evaluation.


1. Novelty Relative to Existing Spherical Operators (Weakness 1)

We thank the reviewer for highlighting Tang et al. (2026) and Snoussi and Karimi (2025). We acknowledge that degree-wise spectral weighting, weight sharing across SH orders, and even-degree SH representations have been investigated previously. We would like to clarify that the main contribution of NFO lies in its CSD-grounded fiber operator formulation and dedicated nonlinear operator network, rather than the spectral multiplication rule alone.

Theoretically, SFNO and GSNO were developed for general spherical function modeling, while Snoussi and Karimi (2025) proposed an equivariant spherical CNN for estimating FODs directly from reduced dMRI measurements. In contrast, NFO specifically formulates LAR-FOD to HAR-FOD reconstruction as a function-to-function mapping on the sphere. Starting from a CSD-motivated fiber system, we derive a Green's-function-based operator representation and explicitly parameterize the learned kernel within the antipodally symmetric, even-degree SH subspace. The spectral computation follows the established SO(3)-equivariant convolution assumption in Eq. (14), yielding the degree-wise parameterization in Eq. (20). This represents the spectral component of NFO, rather than its complete operator formulation.

Architecturally, each NFO block integrates spherical spectral transformations, learnable kernel multiplication (equivariant convolution), 1×1 channel interactions, feature concatenation, and nonlinear activation. Multiple NFO blocks form a complete nonlinear operator network, NFONet, which progressively aggregates neighboring fiber functions to reconstruct the target FOD. This differs from the FOD-Net-based spherical operator adaptations considered in Tang et al. (2026).

We have added the analysis in the revised Related Work and Methodology, with the corresponding changes highlighted in red.

2. Operator and Architecture Contributions (Weakness 2 & Question 2)

We appreciate the suggestion to better isolate the contributions of the proposed operator and network architecture. We have conducted additional experiments at both the complete-network and component levels.

First, we expanded the SOTA comparisons to include SDNet (2024), FOD-SFNO, and FOD-GSNO (2026), in addition to the original FOD-Net and FOD-Swin-Net baselines. FOD-SFNO and FOD-GSNO provide direct comparisons with existing spherical operators evaluated on the same FOD angular super-resolution task.

MethodHCP-YA ACC ↑MGH ACC ↑CC ACC ↑
FOD-Net0.68330.64170.7894
FOD-Swin-Net0.66520.42980.4089
SDNet0.73670.67630.8243
FOD-SFNO0.70690.67030.8165
FOD-GSNO0.71670.67550.8193
NFO0.75330.68270.8333

NFO achieves the highest mean ACC across all three datasets, including the two zero-shot settings. The complete tissue-wise ACC/RMSE results and additional statistics are provided in the revised manuscript (Section 5.1).

Second, we conducted component-wise ablations within the same NFONet architecture to directly assess the contributions of spectral learning, channel interactions, and residual processing.

ModelHCP-YA ACC ↑MGH ACC ↑CC ACC ↑
NFO w/o Spectral Branch0.68910.63180.4956
NFO w/o 1×1 Convolution Branch0.75070.68230.8328
NFO w/o Residual Branch0.74940.67880.8296
NFO0.75330.68270.8333
NFO + GSNO Correction0.75390.68340.8325

Removing the spectral branch decreases ACC by 0.06, 0.05, and 0.34 on HCP-YA, MGH, and CC, respectively, directly demonstrating its substantial contribution to reconstruction accuracy and cross-domain generalization. Removing the residual branch and Removing 1×1 branch also reduce accuracy across all three datasets.

Furthermore, to evaluate an alternative operator parameterization within the same architecture, we introduced NFO + GSNO Correction, incorporating the generalized correction from GSNO while keeping the remaining NFONet unchanged.

Together, these experiments provide complementary evidence for the spectral branch and complete network design. We have added the results and analysis to the revised manuscript (Section 5.1 and 5.5), highlighted in red.

3. Inverse-Operator Assumptions (Weakness 3 & Question 1)

We thank the reviewer for raising this important theoretical concern. We agree that angular undersampling, bandwidth loss, and measurement noise may introduce information loss, and therefore the degradation process does not necessarily admit a unique or exact inverse. We would like to clarify that Eqs. (8)–(9) introduce an idealized, CSD-motivated operator formulation, rather than assuming that the physical degradation process is exactly described by an invertible deterministic linear operator. The inverse expression is conditional on the existence of a suitable inverse. In practice, NFO learns an effective reconstruction mapping between LAR-FOD and HAR-FOD estimates, without requiring exact inversion of the degradation process.

We have clarified this distinction and the corresponding assumptions in the revised Section 4.1 and Appendix A, with all modifications highlighted in red.

4. Clinical Applicability (Weakness 4)

We appreciate the reviewer's concern regarding clinical deployment. We agree that the present experiments do not establish prospective clinical utility or deployment readiness. Our study primarily aims to investigate the feasibility and generalization of spherical operator learning for FOD angular super-resolution.

Beyond reconstruction accuracy, we evaluate downstream tractography and structural connectomes, together with zero-shot generalization across acquisition protocols and from in-vivo to ex-vivo data. These experiments provide complementary evidence of the reconstruction quality and transferability of NFO, but do not replace prospective clinical evaluation.

We have clarified this scope and acknowledged the need for further validation of clinical utility and diagnostic reliability in Appendix B.2 (Limitations) of the revised manuscript, with the corresponding changes highlighted in red.

5. Computational Efficiency (Weakness 5 & Question 3)

We thank the reviewer for highlighting this important trade-off. We agree that parameter efficiency and computational efficiency are distinct properties, and that the inference time of NFO is higher than that of several conventional reconstruction models.

The main computational overhead arises from two factors:

  1. Repeated spherical transformations: Each NFO block performs spectral transformations and nonlinear operations, which introduce additional computational costs beyond the spectral Parametrization.
  2. Voxel-centered inference: NFO predicts the target FOD using a local 9×9×99\times9\times9 neighborhood for each center voxel. In contrast, FOD-Swin-Net employs a patch-wise reconstruction strategy (96×96×9696 \times 96 \times 96), enabling more efficient inference over larger regions.

These differences help explain the substantial runtime gap. However, NFO contains only 2.74M parameters, compared with 19.59M for FOD-Net, 48.17M for FOD-Swin-Net, and 74.93M for SDNet. In particular, compared with the strongest competing method (SDNet), NFO uses approximately 27×27\times fewer parameters and achieves approximately 10×10\times faster inference on HCP-YA, while obtaining higher performance across all three datasets.

Therefore, our intended efficiency claim concerns the accuracy achieved with a relatively compact parameterization, rather than universal inference-speed superiority. We have revised the corresponding statements in the Abstract, Introduction, and Experiments to distinguish these properties and explicitly discuss the accuracy--efficiency trade-off.


We sincerely thank the reviewer again for the valuable and constructive feedback. We have clarified the theoretical positioning and inverse-operator assumptions, added direct spherical operator comparisons and controlled ablations, and strengthened the discussion of computational efficiency and clinical applicability. All corresponding changes are highlighted in red in the revised manuscript.

Dear Reviewer UMJE,

We hope this message finds you well. We have added new operator-level comparisons and component-wise ablations, along with further clarifications of the theory, efficiency and limitation, in response to your constructive comments.

As the discussion phase is nearing its end, we would greatly appreciate any further feedback you may have. Thank you again for your time and consideration.

Best regards,

The Authors

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

This paper studies angular super-resolution of fiber orientation distributions (FODs) in diffusion MRI, where the goal is to predict high-angular-resolution FODs from low-angular-resolution inputs. The authors propose the Neural Fiber Operator (NFO), which reformulates this task from a voxel-wise regression problem into an operator learning problem on the sphere. Specifically, the method learns a mapping from low-resolution to high-resolution representations in spherical harmonic coefficient space and restricts the representation to the even-order spherical harmonic subspace to preserve the antipodal symmetry of FODs. Based on this design, the authors build a patch-based network, NFONet, for whole-brain or partial-brain FOD reconstruction. Experiments on the HCP-YA, MGH, and ex-vivo corpus callosum datasets show that the proposed method outperforms state-of-the-art methods while using fewer parameters.

Critical strengths
  1. The design is clever because it builds the fact that fiber directions have no forward or backward difference directly into the model, so the network cannot produce physically meaningless results. Its novelty is that, instead of treating FOD reconstruction as ordinary 3D image regression, it learns a mapping between spherical fiber functions, which better matches the geometry and physics of brain fibers.

  2. The paper evaluates both in-domain reconstruction and zero-shot generalization across datasets, acquisition protocols, age groups, scanners, and even from in-vivo to ex-vivo data. The authors further evaluate the impact of the proposed method at the levels of tractography and connectome reconstruction. The experiments demonstrate the effectiveness of the proposed method. Moreover, compared with FOD-Net and FOD-Swin-Net, NFO has significantly fewer parameters.

Critical weaknesses
  1. The final derived operator reduces to a per-degree scalar multiplication of even-order spherical harmonic coefficients: g_hat_x^{m,2l} = W_theta(2l) f_hat_x^{m,2l}, where W_theta(2l) depends only on the SH degree 2l and is shared across all orders m. This is a very simple linear and isotropic spectral transform. Since the paper lacks key ablations that remove this spectral branch, the 1×1 convolution, residual connections, and patch aggregation, it is unclear whether the performance gains actually come from the claimed Green’s-function-inspired operator.

  2. The reference high-angular-resolution FODs are generated by CSD-based algorithms rather than direct biological ground truth, so the evaluation may inherit algorithmic biases. As a result, the observed improvements may partly reflect better fitting to a particular CSD output rather than more accurate recovery of true fiber orientations. Additional validation using alternative references or histology-based metrics would strengthen this claim.

  3. Although the model has fewer parameters, Table 7 suggests that it is slower in inference than other methods. The authors provide little discussion of this apparent discrepancy. Besides, only three baseline methods are compared, which seems somewhat limited.

Questions for authors

The final derived operator reduces to a per-degree scalar multiplication of even-order spherical harmonic coefficients: g_hat_x^{m,2l} = W_theta(2l) f_hat_x^{m,2l}, where W_theta(2l) depends only on the SH degree 2l and is shared across all orders m. This operation is essentially a per-degree scaling. Why do the authors refer to it as a Green’s function operator or a neural operator? If this core operator is removed and only the 1×1 convolution, residual connection, and patch aggregation are kept, how much does the performance drop? If such an experiment is provided, I would be willing to raise my score.

5 comments

We sincerely thank the reviewer for the constructive feedback and for recognizing the physical motivation, cross-domain generalization, and parameter efficiency of NFO. We particularly appreciate the suggestions for clarifying the operator contribution and strengthening the experimental evaluation.


1. Operator Formulation and Ablation Studies (Weakness 1 & Question)

We thank the reviewer for this important question. We would like to clarify that the degree-wise spectral multiplication in Eq. (20) represents the spectral parameterization of the proposed Green's-function operator, rather than the complete NFO. Theoretically, starting from the CSD-motivated spherical fiber system, we formulate FOD reconstruction as a function-to-function mapping and derive its Green's-function representation on the antipodally symmetric SH subspace (Section 4.2). The learnable spectral weights correspond to the degree-wise Green's response under the SO(3)-equivariant assumption. Architecturally, each NFO block integrates spherical spectral transformations, learnable kernel multiplication, 1×1 channel interactions, and nonlinear activation. Multiple NFO blocks constitute a complete nonlinear operator network for reconstructing the target FOD from neighboring spherical fiber functions.

To directly evaluate the contribution of these components, we conducted additional ablations under identical training and evaluation settings:

ModelHCP-YA ACC ↑MGH ACC ↑CC ACC ↑
NFO w/o Spectral Branch0.68910.63180.4956
NFO w/o 1×1 Convolution Branch0.75070.68230.8328
NFO w/o Residual Branch0.74940.67880.8296
Full NFO0.75330.68270.8333

The results demonstrate the importance of the spectral branch. Removing it reduces ACC by 0.06, 0.05, and 0.34 on HCP-YA, MGH, and CC, respectively. The particularly large degradation on CC further highlights its contribution to cross-domain reconstruction. Removing the 1×1 branch or the residual branch also reduces ACC, but the effect is smaller. This evidence also proves the significance of the spectral branch.

Regarding patch aggregation, the 9×9×9 patch-learning strategy is adopted from existing method (FOD-Net, SDNet) to incorporate local anatomical context, rather than being a novel component of NFO. Removing it would change the underlying reconstruction setting from patch-based to single-voxel prediction. We therefore focus the ablations on the proposed operator components.

These results have been added to the revised manuscript (Section 4.3 and 5.5).

2. CSD-Derived Reference FODs (Weakness 2)

We appreciate the reviewer's concern regarding the reference FODs.

FOD angular super-resolution is a well-established and practically important task in diffusion MRI, aiming to reconstruct high-quality FODs from acquisitions with fewer diffusion measurements. It has been extensively investigated in previous studies, including FOD-Net, FOD-Swin-Net, and SDNet, providing a well-established setting for evaluating the proposed NFO framework.

We agree that MSMT-CSD-derived reference FODs are algorithmic estimates rather than direct biological ground truth. Nevertheless, they provide a widely adopted and biophysically motivated reference for evaluating FOD reconstruction. Improvements in ACC and RMSE demonstrate better agreement with these high-quality FOD estimates, although they may not fully establish the accuracy of the underlying biological fiber orientations.

To further assess reconstruction quality, our evaluation extends beyond voxel-wise ACC/RMSE to downstream tractography and structural connectome reconstruction, as well as zero-shot evaluation on microscopy-correlated ex-vivo specimens. These complementary analyses provide additional evidence of reconstruction quality and generalization, although independent histological validation remains valuable.

We have clarified this distinction and acknowledged the limitations of CSD-derived references in the revised manuscript (Appendix B.2), with the corresponding changes highlighted in red.

3. Additional Baselines and Inference Efficiency (Weakness 3)

We thank the reviewer for suggesting more comprehensive comparisons. We have expanded the original FOD-Net and FOD-Swin-Net evaluation by including SDNet (2024), FOD-SFNO, and FOD-GSNO (2026). The following table summarizes the mean ACC on held-out HCP-YA and zero-shot MGH/CC datasets:

MethodHCP-YA ACC ↑MGH ACC ↑CC ACC ↑
FOD-Net0.68330.64170.7894
FOD-Swin-Net0.66520.42980.4089
SDNet0.73670.67630.8243
FOD-SFNO0.70690.67030.8165
FOD-GSNO0.71670.67550.8193
NFO0.75330.68270.8333

NFO achieves the highest mean ACC across all three datasets, outperforming the strongest competing method, SDNet, by 0.017, 0.006, and 0.009, respectively. The revised tables and Appendix provide the complete tissue-wise ACC/RMSE results and corresponding statistics.

Regarding inference efficiency, we agree that fewer parameters do not necessarily imply faster inference. NFO incurs additional computational overhead from repeated spherical spectral transformations and voxel-centered processing. Nevertheless, compared with the strongest competing method (SDNet), NFO uses approximately 27× fewer parameters and achieves approximately 10× faster inference on HCP-YA. Although NFO is slower than some other baselines, it provides a favorable reconstruction accuracy–efficiency trade-off.

We have clarified these computational trade-offs and revised the efficiency-related claims accordingly (Section 5.1).


We sincerely thank the reviewer again for the valuable suggestions. We have clarified the operator formulation, provided component-wise ablations and additional baseline comparisons, and strengthened the discussion of CSD-derived references and computational efficiency. All corresponding changes are highlighted in red in the revised manuscript.

Dear Reviewer rJDp,

We hope this message finds you well. We have added new experiments and clarifications in response to your constructive comments, particularly the branch ablation and additional baseline comparisons.

As the discussion phase is nearing its end, we would greatly appreciate any further feedback you may have. Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Rebuttal

Dear Area Chair and Reviewers,

Thank you very much for your time and consideration. We sincerely thank all reviewers for their constructive feedback and valuable suggestions. We have considered the concerns raised in the reviews and substantially strengthened our manuscript. All corresponding changes are highlighted in red in the revised manuscript. We would like to briefly summarize the main improvements and revisions to facilitate further discussion and assessment:

  1. Methodological Novelty: We have clarified the theoretical and architectural distinctions between NFO and existing spherical operators, emphasizing its CSD-motivated fiber operator formulation and dedicated nonlinear multi-block architecture.

  2. Expanded Experimental Comparisons: We have added SDNet, FOD-SFNO, and FOD-GSNO as additional baselines. NFO achieves the highest mean ACC across all three evaluated datasets, including the two zero-shot settings.

  3. Operator and Architecture Ablations: We have added component-wise ablations and a controlled NFO + GSNO Correction experiment, providing further evidence for the contribution of spectral learning and distinguishing operator parameterization from the complete network architecture.

  4. Theoretical Clarifications and Limitations: We have clarified the inverse-operator assumptions, computational trade-offs, limitations of CSD-derived references, and the scope of clinical applicability.

We hope this summary will be helpful for the subsequent discussion and assessment. As the discussion period approaches its conclusion, we would greatly appreciate any further feedback or remaining concerns and would be happy to provide additional clarification if needed. Thank you again for your time and consideration.

Best regards,

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.