List of Papers Browse by Subject Areas Author List
Abstract
Nystagmus is an involuntary rhythmic eye movement that is central to the diagnosis of vertigo and related neurological disorders. It consists of slow phases, reflecting pathological ocular drift, and fast phases, corrective saccades that reset gaze. Quantification of slow-phase velocity (SPV) is clinically essential, providing an objective and interpretable measure for physiological verification of nystagmus dynamics in vestibular assessment. However, traditional SPV estimation typically relies on threshold-based detection and hand-tuned heuristics that can be sensitive to noise, artifacts, and inter-subject variability. We present an end-to-end framework for oculomotor event segmentation that directly supports SPV reconstruction from predicted slow phases. Our key contribution is Harmonic Kernel Mixing, an interpretable spectral front-end that augments a Sinc-parameterized band-pass filterbank with a learnable harmonic cross-band coupling among related frequency bands while retaining transparent cut-off frequencies and coupling weights. Building on this front-end, we introduce a spectro-temporal network with multi-scale temporal encoding to segment slow phase, fast phase, no nystagmus, and artifact regions. On the Omniax clinical dataset with patient-wise stratified 5-fold cross-validation, our method achieves 0.735±0.005 macro-F1, the highest mean macro-F1 among evaluated time-series baselines. We further evaluate on GazeCom and an independent set of 164 BPV cases, where predicted slow-phase segments enable accurate SPV reconstruction (MAE 2.44±2.99 °/s) with <2 s mean absolute error for key SPV temporal landmarks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2470_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/KunalChaturvedi28/NyST-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ChaKun_Harmonic_MICCAI2026,
author = { Chaturvedi, Kunal AND Braytee, Ali AND Yang, Nicholas AND Hannigan, Imelda AND Reid, Nicole AND Bradshaw, Andrew AND Welgampola, Miriam AND Prasad, Mukesh},
title = { { Harmonic Kernel Mixing for Interpretable Oculomotor Event Segmentation in Video-Nystagmography } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a spectro-temporal network for frame-wise oculomotor event segmentation from VNG eye traces, with a particular focus on clinically meaningful slow-phase detection. Its main methodological novelty is Harmonic Kernel Mixing, an interpretable spectral front-end that extends Sinc-parameterized band-pass filtering with learnable harmonic coupling across related frequency bands. Beyond segmentation accuracy, the paper shows that the predicted slow-phase segments can be used to reconstruct SPV profiles and clinically relevant temporal landmarks for BPV assessment.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper addresses a clinically meaningful problem and defines the task at the level of oculomotor event segmentation, which is better aligned with SPV analysis than high-level classification alone. This makes the output more relevant for physiological verification and clinical interpretation.
- The main methodological contribution, Harmonic Kernel Mixing, is original and well motivated.
- The evaluation is solid. The authors use patient-wise 5-fold cross-validation on a sizeable clinical dataset, compare against both classical and deep learning baselines, and also test the method on the public dataset.
- The paper includes clinically relevant verification beyond segmentation metrics. The reconstructed SPV profiles and temporal landmarks on an independent BPV cohort strengthen the claim that the method can support downstream clinical assessment.
- The ablation study is a strength.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The paper discusses recent domain-specific learning-based work for nystagmus analysis, but the experiments mainly compare against generic time-series models.
- The external evaluation is not fully clinical. GazeCom supports cross-dataset testing, but it is not a vertigo/VNG dataset.
- The error analysis is limited, especially for difficult cases such as low-amplitude tails and artifacts.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant problem, the proposed method is reasonably novel and interpretable, and the evaluation includes strong internal testing, external benchmarking, and SPV reconstruction on an independent BPV cohort. However, there is limited comparison with domain-specific baselines, and the fact that the clinical validation remains indirect rather than diagnostic.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal addresses my points reasonably. The authors explain the mismatch between recent VNG diagnosis/detection methods and their segmentation task, clarify the role of GazeCom as an external oculomotor benchmark, and discuss T95 variability and artifact-related errors. I also considered the comments from Reviewers #2 and #3.The rebuttal clarifies important methodological issues, including validation-only hyperparameter selection, patient-wise splitting, annotation protocol, class distribution, and the use of stride-5 windows only for training. Some limitations remain, especially modest gains without statistical significance analysis, indirect clinical validation, and no direct comparison with commercial SPV tools. However, I think the clinical relevance, interpretable HKM front-end, and SPV reconstruction on an independent BPV cohort are sufficient for acceptance.
Review #2
- Please describe the contribution of the paper
The paper proposes NyST-Net, a spectro-temporal model for oculomotor event segmentation in video nystagmography (VNG), designed to enable reconstruction of slow-phase velocity (SPV) profiles for clinical interpretation. The core technical component is Harmonic Kernel Mixing (HKM), which extends Sinc-based band-pass filtering by introducing a learnable harmonic coupling mechanism across frequency bands, aiming to improve robustness while preserving interpretability.
The system performs frame-wise segmentation of eye movement signals into slow phase, fast phase, no nystagmus, and artifact classes, followed by post-hoc SPV reconstruction and extraction of clinically relevant temporal landmarks. The method is evaluated on a private clinical dataset (Omniax) using patient-wise cross-validation and on the public GazeCom dataset. Additional validation is conducted via SPV reconstruction accuracy on an independent cohort of 164 BPV cases.
While the paper is well-motivated from a clinical perspective and demonstrates consistent improvements over selected baselines, the primary contribution lies in a structured modification of existing Sinc-based spectral representations combined with standard temporal modeling (TCN + Transformer), rather than a fundamentally new modeling paradigm.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Focus on physiologically verifiable outputs. The work targets SPV reconstruction rather than only diagnostic classification, improving clinical relevance and alignment with real diagnostic workflows.
2.Appropriate clinical evaluation protocol. The use of patient-wise cross-validation reduces patient-level leakage and reflects realistic generalization.
3.Structured inductive bias in the spectral domain. HKM introduces an interpretable coupling mechanism (coupling matrix) that is more structured than unconstrained channel mixing.
4.Independent clinical cohort evaluation. The inclusion of a separate BPV cohort for SPV error analysis provides stronger end-to-end clinical validation.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Overstated novelty. The method is essentially SincNet-style filtering + constrained channel mixing + standard TCN/Transformer backbone; the “harmonic prior” behaves closer to log-frequency smoothing than true harmonic modeling.
2.Marginal performance gains. Improvements over strong baselines on Omniax are small (~1–2% F1), and without statistical significance analysis the “outperforming” claim is weak.
3.Insufficient methodological transparency. Details on annotation protocol, class distribution, and aggregation under heavy overlapping windows (stride 5) are limited, affecting trust in reported metrics.
4.Incomplete baseline comparison on GazeCom. Important baselines (e.g., gazeNet, REMoDNaV) are not included, and metric comparability is not discussed.
5.Limitations in SPV reconstruction pipeline. Use of absolute velocity removes clinically relevant directional information, and reliance on heuristic post-processing limits robustness and translational value.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a clinically relevant and well-motivated problem, focusing on SPV reconstruction rather than coarse classification, which aligns well with real-world diagnostic workflows in vestibular assessment. The proposed pipeline is coherent and practically oriented, combining segmentation with downstream physiological analysis, and the inclusion of an independent BPV cohort strengthens the clinical perspective of the evaluation.
However, the methodological contribution appears limited in novelty. The proposed approach can be largely interpreted as a combination of a SincNet-based spectral front-end with structured channel mixing and a standard temporal backbone (TCN + Transformer). While the harmonic coupling mechanism is intuitively motivated and introduces some degree of interpretability, it does not constitute a fundamentally new modeling paradigm.
Empirically, the improvements over strong baselines are relatively modest, and the lack of statistical significance analysis makes it difficult to assess the robustness of the reported gains. Additionally, several methodological aspects are insufficiently detailed, including annotation procedures, class imbalance, and the impact of heavily overlapping windows, which raises concerns about the reliability of the evaluation.
From a clinical perspective, although SPV reconstruction is demonstrated, the validation remains indirect, and the absence of downstream diagnostic evaluation limits the strength of translational claims.
Overall, the paper presents a solid and clinically motivated application with a reasonable engineering contribution, but falls short in terms of methodological novelty and depth of evaluation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal satisfactorily addressed my primary methodological concerns regarding patient-wise evaluation, overlap handling, post-processing consistency, and claim wording. The authors also clarified the role of HKM relative to existing Sinc-based approaches and provided a more careful framing of the contribution.
While I still consider the methodological novelty to be moderate and somewhat incremental rather than fundamentally new, I find the work technically sound and clinically relevant. The remaining limitations, including lack of inter-rater agreement analysis and limited external clinical validation, are acknowledged by the authors and do not invalidate the overall contribution.
Review #3
- Please describe the contribution of the paper
The authors present a novel processing pipeline that segments oculomotor events, with a view to further quantifying slow phase velocity for periods of nystagmus. They develop a front-end that ‘softly’ constrains CNN kernels to extract information related to clinically relevant frequencies and their harmonics.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The system developed is shown to perform well compared to the baseline approaches. The system developed is novel in terms of the unique front-end, followed by event classification and an automatic quantification of slow phase velocities. The system is demonstrated to be robust to a range of data aquisition configurations.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
It is assumed that the parameters for NyST-Net, which vary according to the evaluation dataset used, were selected from the validation data. However, this point is not made explicitly and leads to the possibility that the parameters were tuned for the test data.
While it is true that the results of NyST-Net mostly outperform the baseline approaches, sometimes this improvement is marginal and it would be useful to know if these results are statistically significant.
It is unclear what the clinical feasibility of this work is, as there is no report of the processing requirement of the techniques presented. Also, many VOG manufactures use proprietary algorithms for calculating SPV, and it is unclear whether the technique presented here would be superior to such commercially available tools (e.g. https://www.interacoustics.com/balance-testing-equipment/visualeyes)
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
It is possible that the dataset of 164 BPV patients was collected or curated by the authors and such I would have expected to find reference to ethical approvals in this paper.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The results are promising but it’s unclear whether they are statistically significantly better than the baseline approaches. It’s unclear how the approach presented compares in terms of processing time, which is an important factor when it comes to clinical applicability. However, the overall approach is novel and the results suggest that it may offer improved accuracy if evaluated fully.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I do not think that the authors have fully responded to the outstanding question of statistical significance, which was a point raised by all three reviewers. However, after careful consideration, I think this work is worthy of publication.
Author Feedback
We thank the reviewers and AC for the constructive assessment. Methodological details: (AC, R2-Q7,14, R3-Q7): We selected dataset-specific settings using development/validation data only; test folds were not used for hyperparameter selection, model selection, or early stopping. As described in Sec. 3.1, evaluation was patient-wise, with all recordings and 2D components from the same patient kept in the same fold. Regarding annotation protocol, recordings were clinically reviewed before event labels were created by a trained engineer, yielding class distribution: slow phase 54.3%, fast phase 16.2%, no nystagmus 21.2%, and artifact 8.2% across 3,153 traces from 665 patients. Stride-5 windows were used for training only; validation/test used non-overlapping windows for macro- and class-wise F1. Methodological contribution (AC, R2-Q7,14): HKM addresses a limitation of Sinc-based front-ends: although Sinc filters are interpretable band-pass kernels, they treat frequency bands independently, limiting structured cross-band modelling in non-stationary signals. In Sec. 2.1, HKM forms each effective kernel via a learnable ratio-constrained coupling map anchored to compact harmonic ratios, with learned gating controlling cross-band sharing. This preserves cut-off interpretability while adding coupling weights showing how related bands contribute to each effective kernel. Fig. 4 isolates this by fixing the backbone/training and comparing HKM with Sinc-only, unconstrained full mixing, and 1×1 mixing. Experimental comparison (AC, R1-Q7,14, R2-Q7, R3-Q7): We selected baselines through a task-alignment review of VNG literature: methods [7–10] mainly address diagnosis/detection, making them clinically relevant but not directly compatible segmentation baselines. The closest task-aligned comparator is the classical pipeline [11] in Table 1, complemented by representative deep backbones re-implemented under the same protocol. This review also informed public benchmarking: comparable VNG datasets were unavailable, therefore GazeCom was used as an oculomotor benchmark. To support fair interpretation, we will state each method’s source, protocol, and metric definition, including gazeNet and REMoDNaV. For Omniax performance gains, we will align wording with Table 1: NyST-Net achieves the highest mean F1 for the SPV-relevant slow/fast phase labels, with comparable artifact performance. We will revise wordings such as “outperforms” with “highest mean F1” or “class-wise gains.” Clinical interpretation (R1-Q14, R2-Q5): SPV analysis was emphasized because VNG interpretation depends on measurable nystagmus dynamics that clinicians inspect: onset, peak strength, and decay. Quantifying these landmarks requires separating slow phases from fast phases, artifacts, and non-nystagmus periods. NyST-Net targets the segmentation step needed to recover quantitative SPV profiles. The BPV cohort (n=164) in Sec. 3.3 assessed SPV profile recovery from unseen cases, supporting transparent VNG analysis. SPV reconstruction (R2-Q7): We used absolute velocity to evaluate SPV envelopes and landmarks. We agree direction remains clinically important; signed SPV can be computed using the predicted slow-phases for full clinical review. For SPV reconstruction, we fixed LOWESS smoothing and segment filtering before evaluation and applied identically to reference and predicted segments, preventing selective tuning. Error analysis and feasibility (R1-Q7, R3-Q7): In Sec. 3.3, we identified T95 variability as the main error mode, occurring when low-amplitude decay tails were fragmented or artifact segments shifted the 5% decay. For processing requirements, the full inference-to-SPV pipeline took 0.011±0.005 s/recording on the BPV cohort using an NVIDIA L40 GPU. In relation to commercial systems, NyST-Net’s contribution is auditability, linking the reconstructed SPV profiles to reviewable slow-phase/artifact labels. We will clarify these details in the final version.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This submission proposes oculomotor event segmentation method. The reviews are found constructive and converges on major concerns: overstatement (incremental rather than foundamentally new), marginal gains without statistical significance, missing methodological details, missing comparatives (baselines, comparable metrics). Despite the concerns, the method is found clinically relevant, aligning with real-world worksflows. The submission could benefit from clarifications in a rebuttal. For all these reasons, and with respect to the other submissions, the recommendation is towards requesting a rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers has maintained some reservations regarding novelty and lack of statistical significance testing, but also appreciated its technical soundness, positive rebuttal that has clarifyed most evaluation concerns, and has found the work clinically relevant for real-world vestibular diagnostics. Recommendation is Borderline Acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The post-rebuttal landscape remains divided for this submission and even more polarized if any than before the author’s response. Reviewer #2 upgraded to Accept but Reviewer #3 downgraded to Reject. IMHO, a rebuttal that simultaneously moves one reviewer up and one down rather than building consensus does not provide a sufficient basis for acceptance. Further, specific concerns about small sample size per class, patient group overlap, and limited baseline comparison were apparently not resolved.
