List of Papers Browse by Subject Areas Author List
Abstract
Graph networks (GNet) have presented great potential in medical hyperspectral images (MedHSIs) for tumor diagnosis. Most of existing approaches easily suffer from the misconnection of graph inter-nodes due to tissue heterogeneity, imaging noise, and spectral aliasing included in MedHSIs. To remedy such difficulty, a novel active evaluation-induced graph network (AEGNet) based MedHSI classification is proposed for tumor diagnosis. The main innovative contributions are two-fold. (1) Active Evaluation Scheme: We introduce an adversarial discriminator to actively evaluate and refine the correlations (edges) between nodes, enabling more precise graph construction. (2) Hierarchical Calibration Loss: To mitigate the hierarchical misalignment caused by potential inter-node misconceptions, we design a hierarchical calibration loss that enforces consistency between the features of all hidden layers and those of the final layer, improving feature representation across the network. In our experiments, complete empirical evidence has been provided on the In Vivo Hyperspectral Human Brain database, demonstrating that the proposed AEGNet method surpasses several representative approaches in the MedHSI classification based tumor diagnosis task. The ablation discussion has been provided to experimentally reveal the effectiveness of the innovative active evaluation and hierarchical calibration.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2305_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanMei_Active_MICCAI2026,
author = { Wang, Meiling AND Xing, Changda AND Wu, Yifang AND Cao, Liying},
title = { { Active Evaluation-Induced Graph Network Based Medical Hyperspectral Image Classification for Tumor Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes an Active Evaluation-induced Graph Network (AEGNet) for medical hyperspectral image (MedHSI) classification in tumor diagnosis. The method introduces two main components: (1) an adversarial discriminator to dynamically evaluate and refine node correlations (edges) during graph construction, and (2) a hierarchical calibration loss that enforces feature consistency across hidden layers and the final layer to mitigate hierarchical misalignment. Experiments on the In Vivo Hyperspectral Human Brain dataset demonstrate improved performance over several baseline methods, with ablation studies showing the contribution of each proposed component.
- 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.The motivation to address inaccurate inter-node correlation in graph-based MedHSI classification is clinically relevant and technically sound. 2.The adversarial discriminator for dynamic edge evaluation offers a conceptually interesting departure from static graph construction methods. 3.The hierarchical calibration loss provides a mechanism for deep supervision, potentially improving feature robustness across layers.
- 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.The hierarchical calibration loss in Eq. (5) uses Z as the final feature, but Z is not explicitly defined in relation to H(L) in the GNet propagation. It is also unclear whether W(l) is shared across layers or independently learned. 2.The training sample ratio is reported as both 0.05% (Section 3.2) and 0.5% (Figure 2 caption), creating confusion. At 0.05% per class, the total labeled samples are extremely sparse—potentially fewer than 10 per class—which raises concerns about the stability and fairness of comparisons, especially against deep learning baselines that typically require more data. 3.The selected baselines (e. g. , WSGraphCL, GiGCN) are from remote sensing HSI literature. There is no comparison with recent graph-based methods specifically designed for medical hyperspectral data or with standard multi-task/survival models, limiting the contextualization of the claimed improvements. 4.The ablation study isolates components but does not evaluate the effect of the adversarial discriminator without the hierarchical calibration loss in a controlled manner. The current setup only removes one component at a time, leaving potential interaction effects unexplored.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
This paper presents a technically motivated approach to graph-based MedHSI classification with two novel components. However, critical ambiguities in the methodology, significant experimental flaws (including inconsistent training ratios, aggressive spectral downsampling, and lack of statistical rigor).
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper’s main contribution is a graph-based framework for medical hyperspectral image classification that aims to improve tumor diagnosis by making graph construction more reliable under tissue heterogeneity, imaging noise, and spectral aliasing. The main innovative contributions are twofold: (1) an Active Evaluation Scheme, which introduces an adversarial discriminator to actively evaluate and refine the correlations between nodes so that graph edges can be constructed more accurately; and (2) a Hierarchical Calibration Loss, which enforces consistency between intermediate hidden-layer features and the final-layer representation to reduce hierarchical misalignment and improve feature learning across the network. Overall, the work contributes a more robust graph learning pipeline for medical hyperspectral tumor classification.
- 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.The paper addresses a clinically relevant problem. It focuses on tumor diagnosis from medical hyperspectral images, which is a meaningful application area where robust classification methods could be valuable. 2.The proposed method has a reasonably clear technical contribution. In particular, the active evaluation scheme for refining graph edges and the hierarchical calibration loss for aligning intermediate and final features give the framework some methodological originality beyond a standard graph classifier. 3.The paper provides empirical support through baseline comparisons and ablation studies. The experiments show that both the active evaluation component and the hierarchical calibration loss contribute to the final performance.
- 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.The novelty is promising but appears somewhat incremental relative to prior work. The paper builds on an already active line of work in graph-based hyperspectral classification and MedHSI analysis, including methods such as GiGCN, WSGraphCL, DMCA, and FUST, so the manuscript would benefit from more precise clarification of what is fundamentally new beyond combining active edge refinement with hierarchical feature alignment. 2.The evaluation scope is somewhat limited for the strength of the conclusions. Although the In Vivo HHB dataset is described as containing 36 MedHSIs, the main comparisons are presented using only two representative images, making it difficult to assess how consistently the method performs across the full dataset. 3.The train/test protocol could be discussed more carefully. Training and testing are performed using pixels from the same MedHSI, with only a very small proportion used for training, so spatial correlation within the same image may make the task easier than a more realistic cross-image or cross-patient setting. 4.The clinical significance is still somewhat preliminary. The work is relevant to tumor diagnosis, but the validation remains limited to a single dataset without external testing or broader discussion of robustness and clinical deployment, which reduces the evidence for real-world applicability at this stage. 5.The ablation study is helpful but still fairly narrow. It shows that the two proposed components contribute to performance, but it would be more convincing to also compare against simpler graph-construction strategies or stronger graph-based baselines to better isolate the sources of the gains. 6.Some parts of the method description would benefit from clearer explanation. In particular, the optimization procedure is summarized only briefly, and the role of the “adversarial discriminator” is not fully intuitive from the current presentation, so reproducibility and interpretation are somewhat limited. 7.The manuscript would also benefit from language polishing. There are several grammatical issues, typos, and inconsistencies that affect clarity, including phrases such as “inter-node misconceptions,” “the correlation between nodes is true,” the mismatch between 0.05% in the text and 0.5% in the figure captions, and wording such as “AEGNet form without…”.
- 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 submission does not provide sufficient information for reproducibility.
- 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
The paper attempts to solve a relevant problem and includes some interesting ideas, particularly the active evaluation mechanism and the hierarchical calibration loss. To strengthen the work, it would be helpful to clarify the methodological novelty more explicitly relative to prior graph-based hyperspectral classification methods, expand the evaluation beyond a small number of representative images, and discuss the limitations of the current same-image train/test setting more carefully. The manuscript would also benefit from language polishing, since several grammatical issues, wording inconsistencies, and minor typos currently make parts of the method and experimental setup harder to follow.
- 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?
My recommendation of weak reject is mainly because the paper has a relevant application and some interesting technical ideas, but the current evidence is not quite strong enough for acceptance. In particular, the active evaluation mechanism and hierarchical calibration loss are reasonable contributions, and the reported results are encouraging. However, the overall novelty feels somewhat incremental relative to prior work on graph-based hyperspectral classification and MedHSI, and the paper does not fully clarify how large the methodological step is beyond existing directions. The main factors that lowered the score were the limited scope of the evaluation and the unclear strength of the validation protocol. Although the dataset is described as containing 36 MedHSIs, the main comparisons are presented using only two representative images, and the training/testing setup is based on pixels from the same MedHSI, making it difficult to assess robustness in a more realistic cross-image or cross-patient setting. In addition, the clinical significance remains somewhat preliminary because there is no external validation or broader discussion of generalizability. Another factor was clarity. Some parts of the method, especially the optimization description and the role of the so-called adversarial discriminator, would benefit from clearer explanation, and the manuscript also contains noticeable grammatical issues and inconsistencies that make it harder to follow. Overall, I see the work as marginally below the acceptance threshold at this stage, but I would be open to revising the assessment if the rebuttal better justifies the novelty, clarifies the experimental design, and addresses the validation concerns.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors’ response helps clarify several points, particularly the correction of the training ratio, the role of the final-layer feature in the hierarchical calibration loss, and the intended function of the discriminator within the optimisation process. The additional discussion of alternative graph-construction strategies is also helpful. That said, I would still keep my recommendation at Weak Reject, because the main concerns are only partially resolved. In particular, the central evaluation in the submitted paper remains limited: the reported comparisons are based primarily on two representative MedHSIs from a larger dataset of 36 images, and the within-image train/test protocol makes it difficult to assess robustness in more realistic cross-image or cross-patient settings. While the rebuttal outlines additional experiments planned for the camera-ready version, these do not fully address the evidence presented in the current submission. I also still find the methodological advance somewhat incremental relative to existing graph-based hyperspectral classification work. Overall, the rebuttal improves the paper and addresses some clarity issues, but not enough for me to move the paper above the acceptance threshold.
Review #3
- Please describe the contribution of the paper
The paper introduces AEGNet, a novel graph-based framework for medical hyperspectral image (MedHSI) classification. Its primary contribution is the integration of an adversarial discriminator that actively evaluates node correlations to refine graph topology, moving beyond static spatial or spectral thresholding. Additionally, it proposes a hierarchical calibration loss to ensure semantic consistency across network layers, addressing the common GNN issue of feature misalignment.
- 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.The use of an adversarial scheme to dynamically prune “noisy” edges caused by tissue heterogeneity is a technically sound approach to MedHSI analysis. 2.The hierarchical calibration mechanism acts as a form of deep supervision, which potentially mitigates over-smoothing and stabilizes feature distributions in deep graph architectures. 3.Empirical results on the In Vivo Human Brain database show that the model achieves high accuracy (e.g., 98.28% OA on specific samples), outperforming several state-of-the-art baselines like WSGraphCL and GiGCN. 4.The work targets intraoperative tumor diagnosis, a high-impact clinical application where real-time spectral-spatial analysis is crucial.
- 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.The validation is limited to a small subset of the In Vivo HHB dataset (randomly selecting only two MedHSIs for detailed comparison). This raises concerns regarding the generalizability of the results across a broader patient population. 2.The text mentions using 0.05% of the sample as training. However, there is another 0.5% in the illustration of Figure 2, which is a bit contradictory, please check the author. 3.The paper says that the spectral band is reduced from 826 to 27, and although such a large dimensionality reduction can simplify the calculation, a large amount of spectral information will be lost. Does this practice have a significant impact on graph structure learning? It is recommended to add comparative experiments with different dimensionality reduction scales.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
I recommend a Weak Accept. The technical motivation for using adversarial evaluation and hierarchical calibration to handle the complexities of MedHSI data is compelling and shows promising results. While the limited dataset scale are notable issues, the core methodology offers a meaningful incremental advancement for the MICCAI community. If the authors rectify the typographical errors and provide a more comprehensive statistical evaluation (including clinical metrics and efficiency analysis), this work could be a valuable contribution to medical graph representation learning.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We would like to express our sincere gratitude to the reviewers and address your main concerns below.
<Clarification on training sample ratio(@R1 & R2 & R3)> The training ratio is 0.5%. The writing error in figure caption will be corrected into 0.5%.
<Explanation on hierarchical calibration loss(@R1)> We state that Z = H(L) in Section 2.1 (Hierarchical Calibration), where H(L) is the output feature of the last GNet layer. That is, Z is the final-layer feature of the graph network, used as the target for hierarchical alignment. Also, each transformation matrix W(l) is independently learned for each hidden layer.
<Approach comparison(@R1)> Due to the similar imaging mechanisms, our initial baselines included widely adopted remote sensing HSI methods as they are commonly compared in medical HSI studies. We will further add comparisons with state-of-the-art medical-specific graph methods in the camera-ready manuscript to better contextualize our improvements.
<Clarification on ablation study(@R1)> Actually, we have initially evaluated the effect of the adversarial discriminator without the hierarchical calibration loss, with the results shown in Fig.4(b), where the AEGNet_NH is the AEGNet with the adversarial discriminator and without the hierarchical calibration.
<Clarification of novelty(@R2)> Unlike existing graph-based methods that suffer from severe node misconnection, our AEGNet innovatively introduces adversarial active edge evaluation to dynamically refine graph edges, and further embeds hierarchical calibration to stabilize feature learning. The method fundamentally solves the challenge of inaccurate graph modeling in hyperspectral tumor diagnosis.
<Evaluation scope(@R2)> It is common for existing hyperspectral work to use several representative images for evaluation like our original manuscript. Following your suggestion, we have added quantitative results on two additional MedHSIs in the camera-ready version to further verify the consistent performance of our method on the In Vivo HHB dataset.
<Train/test protocol(@R2 & R3)> Our experiments follow the within-image train/test setting (splitting pixels of the same MedHSI into training and testing sets) with an extremely low ratio of training samples, which is a widely adopted and fair protocol in medical hyperspectral image classification literature. To address your concern, we have added a detailed discussion on the cross-image evaluation in the camera-ready version.
<Clinical significance(@R2)> Our validation has been currently performed on a single clinical dataset. According to your advice, we will add the experimental validation on the Cholangiocarcinoma Recorded Multidimensional Choledoch (MC) dataset.
<Narrow ablation study(@R2)> We have compared several graph construction strategies including fixed threshold, k-nearest neighbor, and spatial-only graph, with the results will be shown in the camera-ready version. It’s found that the AEGNet (OA=95.30) achieves the higher value that the forms with fixed threshold (OA=91.01), k-nearest neighbor (OA=93.64), and spatial-only graph (OA=93.15).
<Optimization procedure(@R2)> The adversarial discriminator is trained via binary cross-entropy loss to dynamically evaluate correlations between nodes, distinguishing intra-class and inter-class pairs. Its output directly defines the edge weights in the graph, which guides graph construction throughout the entire alternating optimization process. We will show the clearer optimization procedure in the camera-ready version.
<Impact of spectral band(@R3)> Initially, we used the 27-band setting to clearly highlight performance differences across methods. Further, we have conducted comparative experiments with 100, 50, and 27 bands, and the results will be shown in the camera-ready version. The results show that our method maintains consistent performance across different scales, validating that the graph learning is not negatively impacted by the reduction to 27 bands.
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.
The paper introduces an Active Evaluation-Induced Hyperspectral Image Classification Graph Network for tumor diagnosis. Reviewers R1 and R3 consider the technical motivation compelling and the methodology meaningful, although R1 highlights notable ambiguities and weaknesses in the experimental design. R2 views the contribution as somewhat incremental, with a limited evaluation scope and several aspects requiring further clarification.
- 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 authors’ rebuttal addresses the main methodological ambiguities and some experimental setup concerns raised by the reviewers, including clarifying the training ratio, the definition of the final-layer feature in the hierarchical calibration loss, and the role of the adversarial discriminator in the optimization. They also outline plans to extend validation to another clinical dataset, which together strengthen the empirical support for the method. Given these clarifications, the clinically relevant application, and the fact that two reviewers remain at weak accept after rebuttal, I consider the remaining limitations acceptable for MICCAI.
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.
While the paper received two positive reviews (who did not provide final ratings after the rebuttal), one reviewer gave a reject because of insufficient novelty and experimental evaluation. The reported comparisons are based primarily on only two representative MedHSIs from a larger dataset of 36 images. Moreover, the paper has formatting issues. The spacing around section titles has been significantly reduced and figure captions are a bit smaller than the standard font.
Meta-review #3
- 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 rebuttal clarifies several implementation ambiguities, but the main validation concern remains insufficiently resolved because the submitted evaluation is still mainly based on a within-image pixel split and only a small number of representative MedHSIs. The proposed adversarial edge refinement and hierarchical calibration are reasonable, but the novelty is incremental and the evidence for cross-image, cross-patient, or broader clinical generalization is not strong enough to support acceptance.