List of Papers Browse by Subject Areas Author List
Abstract
Medical Hyperspectral Imaging (MHSI) provides rich pixel-level biochemical information and has demonstrated strong potential in computational pathology. However, the high spectral dimensionality and strong inter-band redundancy make effective spectral–spatial representation learning challenging. Most existing methods encode spectra as flat channel-wise features or tokenized sequences, without explicitly modeling their intrinsic multi-scale structure. To address this limitation, we propose a wavelet-guided spectral–spatial segmentation framework for MHSI. By performing one-dimensional wavelet decomposition along the spectral dimension, we explicitly separate global biochemical trends and local spectral variations. Instead of replacing the original spectrum, the low-frequency component is incorporated as a structural prior via a per-pixel adaptive scalar gate, stabilizing deep spectral learning. Furthermore, we introduce Band-Aware Aggregation and Spectral-Enhanced Skip Connections to enhance spectral–spatial interaction in the decoding process. Extensive experiments on public medical hyperspectral datasets demonstrate that our method consistently improves segmentation performance compared with state-of-the-art methods. The source code is publicly available at https://github.com/Junya99/Wavelet-Guided-MHSI.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5158_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Junya99/Wavelet-Guided-MHSI
Link to the Dataset(s)
MDC dataset: https://www.kaggle.com/datasets/hfutybx/mhsi-choledoch-dataset-preprocessed-dataset
MOD dataset: https://drive.google.com/drive/folders/124NCpp3-8DKHOai_4nV0ytsUm78sfKzj
BibTex
@InProceedings{JiJun_WaveletGuided_MICCAI2026,
author = { Ji, Junya AND Stone, Nick AND Meng, Yanda AND Ye, Xujiong},
title = { { Wavelet-Guided Spectral–Spatial Learning for Medical Hyperspectral Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes a novel wavelet-guided framework for spectral-spatial segmentation of medical hyperspectral images. Instead of using the wavelet transform as a preprocessing step, the model reparameterizes spectral representations in the wavelet domain and incorporates a gated low-frequency structural prior for spectral modeling.
- 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 is highly relevant, is well-written and easy to follow.
- The proposed pipeline is evaluated against the state-of-the-art on two different medical hyperspectral image segmentation datasets through quantitative benchmarking and shows superior performance.
- A detailed ablation study reveals the contribution of the different components of the design to the performance of the framework.
- 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 following are not major weaknesses as such, but are still worth addressing.
- The font size of the text in the figures is too small and makes interpretation on A4/letter paper (almost) impossible.
- The “additional boundary supervision term” is mentioned for the training loss without further specification. A clarification would be appropriate.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is well-motivated, is detailed and clearly written, and presents convincing experimental results. I think this is a valuable contribution to the field.
- 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 authors convincingly addresses major concerns in their rebuttal.
Review #2
- Please describe the contribution of the paper
The paper proposes a wavelet-guided spectral-spatial framework for medical hyperspectral image segmentation. Its main contribution is to use low-frequency wavelet components as an adaptive structural prior for spectral modeling, together with Band-Aware Aggregation and Spectral-Enhanced Skip Connections to improve spectral-spatial interaction during decoding.
- 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 an important representation issue in medical hyperspectral image segmentation, namely that spectral signatures are multi-scale but are often modeled as flat channel-wise features or tokenized sequences. This motivation is clear and relevant.
- The wavelet-guided design is interesting. Using the low-frequency wavelet component as an adaptive structural prior, rather than as a simple preprocessing step or full replacement of the spectrum, is a reasonable and technically meaningful contribution.
- The experimental results are solid. The method is evaluated on two public medical hyperspectral datasets and shows consistent improvement over strong CNN, Mamba and Transformer-based baselines.
- The ablation study is a strength. It supports the choice of residual LF integration, the db2 basis, and the proposed decoding modules, which makes the paper more convincing.
- 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 methodological novelty is moderate. The paper combines wavelet decomposition, attention-based spectral modeling, and U-Net-like decoding modules, and the main novelty is in how these components are integrated rather than in a fundamentally new formulation. The paper itself cites prior wavelet-based MHSI work and recent wavelet-guided hyperspectral deep models.
- The empirical gains are relatively modest. This is especially true on the MOD dataset, where the improvement over the strongest baseline is small.
- The paper does not provide enough analysis of computational cost or scalability. Since medical hyperspectral data are high-dimensional, runtime, memory use, and comparison of efficiency would strengthen the practical value of the method.
- 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 is technically sound, well organized, and shows improvement on medical hyperspectral segmentation benchmarks with a reasonable methodological contribution. However, the novelty is moderate, the gains are not large, and the clinical validation remains limited.
- 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 clarify that the novelty is the use of the LF component as a pixel-adaptive gated structural prior inside the network. They also explain the smaller gain on MOD by its lower spectral richness, and provide useful efficiency clarifications, including parameter count and the small runtime contribution of the 1D wavelet operation. The rebuttal addresses Reviewer #1’s questions about figure readability and the boundary supervision term, and it addresses Reviewer #3’s concerns about fair baseline retraining and inference latency. Some limitations remain, especially the incremental nature of the contribution and limited clinical validation. However, the method is technically sound, the evaluation on MDC and MOD is solid, and the ablation study supports the roles of residual LF guidance, db2, BAA, and SESC.
Review #3
- Please describe the contribution of the paper
The paper introduces a method to decompose hyperspectral signatures using 1D wavelet transform, separating low-frequency (global biochemical trends) and high-frequency (local spectral variations), rather than treating spectra as flat channels or tokens. The obtained low-frequency part is used as a per-pixel gated structural prior to guide and stabilize deep spectral feature learning. In addition, Spectral-Enhanced Skip Connections and Band-Aware Aggregation modules have been explored to improve how spectral information is propagated and fused, enabling adaptive band-wise feature reweighting during reconstruction. Experiments on public datasets demonstrate that the proposed method improves segmentation performance compared with SoTA methods.
- 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 proposes to use 1D wavelet transform to decompose hyperspectral signatures instead of treating spectra as flat tokens, aiming to physically separate global biochemical trends from local variations, mimicking professional signal processing.
The employment of low-frequency components as a per-pixel gated structural prior potentially anchors the deep learning process, providing stability and reducing noise interference.
The used Spectral-Enhanced Skip Connections and Band-Aware Aggregation possibely addresses the specific challenges of hyperspectral data, such as inter-band redundancy and the need for adaptive feature reweighting.
The proposed model outperforms State-of-the-Art (SoTA) benchmarks, proving that modeling the intrinsic multi-scale structure of spectra yields measurable performance gains.
- 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.
While the paper frames the use of 1D wavelet transforms as a key contribution, the decomposition of hyperspectral signatures into multi-scale components (low vs. high frequency) is a long-standing technique in the field.
The paper claims superiority over State-of-the-Art (SoTA) methods, but it leaves several critical questions unanswered regarding the experimental setup. There is no indication if the SoTA models were retrained from scratch using the same hardware, optimizer, and data splits. If the authors compared their results against the reported numbers in original papers, the comparison is likely invalid due to differences in preprocessing, spectral calibration, and augmentation strategies. It is unclear if the SoTA models were given the same level of hyperparameter optimization as the proposed method, which often leads to baseline bias.
The introduction of specialized modules like Band-Aware Aggregation (BAA) and Spectral-Enhanced Skip Connections (SESC) inevitably increases the network’s parameter count and operations. Without reporting MACs, Parameters (M), and Throughput (FPS), it is impossible to determine if the performance gains are due to the wavelet-guided intelligence or simply because the model is significantly larger/deeper than the SoTA baselines.
For medical applications (especially real-time pathology), latency is a critical constraint. The 1D wavelet transform adds a preprocessing step that, while mathematically simple, must be accounted for in the total inference pipeline.
- Please rate the clarity and organization of this paper
Satisfactory
- 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 presents a technically sound wavelet-guided framework for MHSI segmentation, effectively utilizing 1D wavelet decomposition to separate global biochemical trends from local spectral noise. The inclusion of specialized modules like Band-Aware Aggregation demonstrates a sophisticated approach to handling inter-band redundancy.
However, the technical novelty is incremental, wavelet transforms are long-established in HSI, and the evaluation lacks critical transparency. Specifically, the absence of computational cost metrics (MACs/throughput) and confirmation of fair baseline retraining leaves the practical clinical utility unverified.
- 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.
While the technical novelty is somewhat incremental, the paper presents a solid and well-designed framework for MHSI segmentation with effective integration of wavelet decomposition and band-aware modeling. Overall, the work provides meaningful empirical and application value to the medical HSI community.
Author Feedback
We sincerely thank all reviewers and the AC for their constructive and detailed feedback. We appreciate the following encouraging comments: “well-written and easy to follow…convincing experimental results” [R1]; “The wavelet-guided design is interesting…is a reasonable and technically meaningful contribution” [R2]; “sophisticated approach to handling inter-band redundancy” [R3].
R1Q1 (Clarity) A: We acknowledge the readability issue and will increase the font size of all figures in the revised version to ensure clarity.
R1Q2 (Boundary supervision term) A: The boundary supervision term is an auxiliary loss applied to the predicted edge map and combined with the segmentation loss. The boundary target is automatically generated from the ground-truth mask and requires no additional annotations.
R2Q1 & R3Q1 (Novelty) A: We respectfully clarify that we do not claim the wavelet decomposition itself as the novelty. The key contribution is how the LF spectral component is used inside the network. Here, we integrate the LF component into the network as a pixel-adaptive gated structural prior for spectral representation learning, which is significantly different from prior MHSI wavelet work only used as external preprocessing for denoising or compression. The ablation study (Table 2) further suggests that this improvement does not arise from a naive combination of components, since directly using LF features or simple LF+HF fusion yields weaker performance than the proposed residual LF guidance design. The final improvement comes from the specific residual LF guidance together with decoder-side spectral propagation through BAA and SESC. Thus, the novelty lies in spectral representation learning for clinical MHSI segmentation.
R2Q2 (Performance on the MOD) A: We agree that the gain on MOD is relatively modest compared to MDC, which is reasonable given their characteristics: MOD contains only 16 spectral bands and 138 imaging scenes, while MDC has 60 spectral bands and 538 scenes. Since our method is designed to exploit structured spectral information, the advantage is naturally more visible on the richer MDC dataset. Nevertheless, the method still obtains the best IoU, DSC, and HD on MOD (Table 1), indicating robustness of the method rather than dataset-specific overfitting. To our knowledge, this is also one of the first dense segmentation studies on the recently released MOD dataset (2024).
R2Q3 & R3Q3 (Efficiency and scalability) A: Although explicit efficiency metrics were not reported, Table 2 (row 6) already shows that the proposed framework outperforms strong baselines even before BAA and SESC are introduced, indicating that the primary gain comes from the wavelet-guided LF prior rather than increased model complexity alone. In fact, the proposed model has slightly fewer parameters (26.61M) than the original baseline (FSS, 27.06M), since the heavier low-rank module is replaced by lightweight feature refinement modules (BAA/SESC) operating on grouped features and skip pathways, rather than introducing additional heavy global modeling modules.
R3Q2 (Fair baseline comparison) A: We’d like to confirm that all compared methods were re-implemented and retrained from scratch under an identical experimental protocol, including the same data splits, preprocessing pipeline, and data augmentation strategies. All methods were given comparable hyperparameter tuning effort to ensure a fair comparison. All trained model weights are stored, and the reported results can be consistently reproduced under our implementation.
R3Q4 (Inference latency) A: We’d like to clarify that the wavelet transform in our framework is not an external preprocessing step, but an internal differentiable operation integrated into the network forward pass. Runtime profiling shows that the proposed 1D wavelet operation contributes only 1.1% of the total forward time. The transform itself uses fixed spectral filters and introduces negligible additional overhead.
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 a wavelet-based segmentation of MHI images. The reviews are found constructive and converges on concerns on moderate novelty (integrating established components), on modest gains to chosen baselines, serious readibility issues (sizes in figures), but note a solid validation. A rebuttal should better situate novelty (r23), performance gain (r23), clarify clinical feasibility (r3), as well as address necessary clarity issues in the methodology (r123). For these reasons, 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 reached a consensus on recommending acceptance. The rebuttal has resolved the initial concerns about novelty and computational efficiency.
Meta-review #2
- 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.
While the technical novelty is somewhat incremental, the paper presents a solid and well-designed framework. All reviewers recommended acceptance.
Meta-review #3
- 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.
All three reviewers now recommend acceptance after rebuttal. The rebuttal effectively clarifies the core novelty, addresses efficiency and fairness concerns, and resolves minor clarity issues. Some limitations remain, mainly the incremental nature of the contribution and limited clinical validation, but the method is technically sound, empirically well supported, and of clear interest to the medical HSI community.