List of Papers Browse by Subject Areas Author List
Abstract
Pixel-level annotation for skin lesion segmentation is costly and subjective. This limits large-scale clinical deployment. Weakly supervised methods based on image-level labels alleviate this burden. However, most existing methods rely on class activation maps (CAMs), which often fail to cover the entire lesion. To address this, we introduce a Frequency and Geometry Guided Graph Clustering framework, reformulating weakly supervised skin lesion segmentation as a differentiable node-clustering problem. First, a Frequency-aware Feature Modulation module injects local spectral priors into deep features. This enhances the model’s sensitivity to ambiguous boundaries. Second, we construct a semantic-spatial graph from these features. A graph neural network (GNN) then captures long-range dependencies to ensure structural completeness. Finally, we propose a tri-source fusion strategy that integrates three cues (GNN cluster, CAM, and the input image) via an improved side-window mean filtering. Experiments on ISIC 2017, ISIC 2018, and PH$^2$ show that our method achieves state-of-the-art performance. It significantly narrows the gap with fully supervised models, providing a reliable solution for label-efficient skin lesion analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5172_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hualei213/FG3-Cluster
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DenZha_Frequency_MICCAI2026,
author = { Deng, Zhaoxin AND Wang, Dongang AND Shen, Hualei},
title = { { Frequency and Geometry Guided Graph Clustering for Weakly Supervised Skin Lesion Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
1.A Novel Graph-Clustering Paradigm for Weakly Supervised Modeling: This paper reformulates weakly supervised skin lesion segmentation as a differentiable graph node-clustering problem (DGC module). By utilizing Graph Neural Networks to explicitly model the global topological structure of the lesion area, the framework effectively mitigates the fragmentation issues commonly observed in traditional Class Activation Map (CAM) methods, which tend to focus solely on the most discriminative local regions. 2.Incorporation and Fusion of Frequency-Domain Priors: A Frequency-aware Feature Modulation (FFM) module is proposed. It modulates deep features by extracting the local spectral energy distribution of image patches, thereby enhancing the model’s discriminative capability against complex textures such as hair occlusions and inflammatory boundaries, addressing the limitations of purely spatial domain modeling. 3.Geometry-Constrained Pseudo-Label Refinement: A Tri-Source Refinement (TSR) strategy is designed. It integrates graph clustering predictions, CAM responses, and edge constraints derived from the input image via an improved Side Window Mean Filter (SWMF). This improves the geometric consistency and structural completeness of pseudo-labels along lesion boundaries. 4.Comprehensive Experimental Validation: The method achieves state-of-the-art performance on three public benchmarks (ISIC 2017, ISIC 2018, and PH²). Extensive ablation studies confirm the effectiveness of each proposed module, demonstrating a significant reduction in the segmentation accuracy gap between weakly supervised and fully supervised 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.
1.Holistic Integration: The paper successfully integrates three distinct sources of information—graph topology, frequency domain, and image geometry—into a unified framework. The combination addresses the common WSSS pitfall of fragmented CAMs in a principled manner. 2.Task-Appropriate Design (FFM): The use of patch-wise FFT energy bands to combat dermoscopic artifacts (hair, bubbles) is well-motivated. The FiLM-based modulation is a clean implementation that directly targets the texture ambiguity problem in skin imaging. 3.Clear Presentation: The paper is generally well-written, and Figure 1 provides a comprehensive visual overview of the pipeline. The mathematical formulation is detailed and largely reproducible. 4.Strong Ablation Insights: The stepwise ablation (Baseline -> S1 -> S2 -> S3 -> Ours) effectively isolates the impact of clustering versus frequency priors versus geometric refinement.
- 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 framework is well-engineered, the review identifies several significant concerns regarding incremental novelty, missing implementation details critical for clinical translation, and insufficient comparison with the current state-of-the-art in the While the overall framework is engineered with clear logic and the empirical gains over CNN-based baselines are evident, several significant concerns regarding incremental novelty, incomplete validation against the current state-of-the-art, and missing practical details for clinical translation collectively limit the contribution of this submission.
A primary concern is the incremental nature of the individual technical components. The Frequency-aware Feature Modulation (FFM) module, although effective in the ablation study, applies well-established spectral analysis techniques to this specific domain. The use of patch-wise Fast Fourier Transform (FFT) energy bands to capture texture artifacts such as hair occlusions is a sensible adaptation, but it constitutes an application of known frequency rectification principles (e. g. , as explored in FFR [25]) rather than a fundamental methodological innovation in frequency-guided network design. Similarly, the Tri-Source Refinement (TSR) module relies heavily on the Side Window Mean Filter [26], with modifications to the scoring function (Eq. 9) that are standard variance-based edge-preserving heuristics. As a training-free post-processing step, TSR provides a practical boundary cleanup but does not represent a core algorithmic advance, and the paper lacks a comparison with simpler, widely-adopted refinement alternatives such as DenseCRF to demonstrate the specific necessity of the proposed filter design.
Furthermore, the experimental validation does not adequately position the work within the current state of weakly supervised semantic segmentation (WSSS). The paper evaluates against several strong CNN-based baselines (e. g. , MoRe [24], ToCo [19]), yet the field has seen a decisive shift toward Vision Transformers (ViTs), whose inherent self-attention mechanisms are designed to capture the very long-range dependencies that the proposed Graph Neural Network (GNN) clustering module aims to address. The absence of any comparison with a ViT-based WSSS backbone (such as AFA [18], which is cited but not benchmarked, or other contemporary methods) weakens the claim that the added complexity of differentiable graph clustering is necessary rather than compensatory for the limitations of the ResNet34 CNN architecture. Without such a comparison, it is difficult to assess whether the gains stem from the GNN framework itself or from addressing the well-known inductive bias limitations of CNNs.
Additional weaknesses pertain to practical deployment considerations and reproducibility. The paper lacks a discussion of computational efficiency, reporting training on a high-end GPU but omitting inference latency and memory footprint metrics. Given the clinical application context, understanding whether the graph construction and GNN forward pass scale gracefully to standard dermoscopic resolutions (e. g. , 512x512) is important for assessing translational potential. Moreover, the description of label extraction for the ISIC 2018 dataset is vague; the statement that image-level labels were “extracted from diagnostic metadata” lacks sufficient detail for independent reproduction, which is particularly concerning given that the largest performance margins are reported on this dataset. Finally, the framework introduces several new hyperparameters—most notably the number of clusters M M in the DGC module—without any sensitivity analysis or justification for the chosen values. In a binary segmentation setting, the optimal M M is not obvious, and the robustness of the method to this design choice remains unexplored, leaving open questions about the generalizability of the reported results across varying lesion complexities.
- 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
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?
My recommendation of Weak Reject is primarily driven by a combination of limited methodological novelty, insufficient benchmarking against the contemporary state-of-the-art, and the absence of practical deployment considerations that are expected for a clinical image computing venue such as MICCAI.
The paper presents a well-integrated framework that combines graph clustering, frequency-domain modulation, and geometric refinement for weakly supervised skin lesion segmentation. The individual components are logically motivated, and the ablation study convincingly demonstrates that each module contributes positively to the final pseudo-mask quality. However, the technical novelty of these components is largely incremental. The Frequency-aware Feature Modulation module applies established spectral analysis techniques, and the Tri-Source Refinement module relies on a modest adaptation of existing side-window filtering. While the integration is thoughtful, the core methodological advances do not rise to the level of significant innovation expected for acceptance.
More critically, the experimental validation does not adequately contextualize the contribution within the current landscape of weakly supervised semantic segmentation. The baseline comparisons are restricted to CNN-based methods, whereas the field has increasingly adopted Vision Transformers, whose self-attention mechanisms naturally address the long-range dependency problem that the proposed graph neural network aims to solve. Without a comparison against a ViT-based counterpart, it is difficult to determine whether the performance gains are attributable to the GNN formulation or simply to mitigating the known limitations of convolutional backbones. This omission substantially weakens the claim of state-of-the-art performance.
Additional factors contributing to the decision include the lack of discussion on computational efficiency metrics, which are relevant for assessing clinical translatability, and the insufficient detail provided regarding label extraction for the ISIC 2018 dataset, which raises reproducibility concerns. The absence of sensitivity analyses for key hyperparameters, particularly the number of graph clusters, further limits confidence in the generalizability of the reported results.
Taken together, while the work is technically sound and demonstrates clear improvements over older CNN-based baselines, the marginal novelty and incomplete benchmarking relative to modern WSSS approaches place it below the acceptance threshold for MICCAI. I believe the paper would be better suited for revision and resubmission to a journal venue where the authors can address these limitations with additional experiments and more comprehensive comparisons.
- 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.
After careful consideration of the rebuttal and the other reviews, I maintain my recommendation of reject. The authors failed to effectively address the key concerns; the overall contribution remains an engineering integration of established techniques and does not meet MICCAI’s expectations for methodological novelty and empirical rigor.
Review #2
- Please describe the contribution of the paper
The paper presents a Frequency and Geometry Graph Guided Clustering framework for weakly supervised skin lesion segmentation. The motivation and presentation of the paper are clear. The design and connection chain of the DGC and TSR modules appear to be reasonable. Furthermore, the experiments show that the method is effective, making the paper relatively novel overall. However, my major concern lies in the FFM module. There might be a granularity mismatch issue. Specifically, the FFM module extracts highly localized, fine-grained patch frequency features, but these features are supervised solely by a global image-level label. It is questionable whether such coarse-grained supervision can effectively drive the learning of localized frequency modulations without feature degeneration.
- 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 overall framework is clearly presented and easy to follow. 2.Integrating frequency with geometric graph for weakly-supervised skin lesion segmentation is a reasonable and relatively novel attempt. 3.The experimental results illustrate that the proposed method achieves promising 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.
In the design of the FFM module, the image is divided into patches, and the frequency energy is calculated across different frequency bands. Thus, F_mod represents features containing patch-level frequency energy. Supervising such local frequency information with a global image-level label may cause a granularity mismatch issue. The method appears to be primarily an integration of well-established techniques (e.g., FFT, FiLM, and standard GNN clustering). The author could be better clarify the deeper effect between frequency domain and geometric graph.
- 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?
In the design of the FFM module, the image is divided into patches, and the frequency energy is calculated across different frequency bands. Thus, F_mod represents features containing patch-level frequency energy. Supervising such local frequency information with a global image-level label may cause a granularity mismatch issue.
- 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 #3
- Please describe the contribution of the paper
Addressing the challenge in weakly supervised skin lesion segmentation—where Class Activation Maps (CAMs) often cover only the most discriminative local regions and struggle to recover the complete lesion structure—this paper proposes a Frequency and Geometry Guided Graph Clustering (FG3-Cluster) framework. The proposed method reformulates weakly supervised segmentation as a differentiable graph node clustering problem and designs three core modules: Frequency-aware Feature Modulation (FFM) to inject frequency-domain priors; Differentiable Graph Clustering (DGC) to model long-range dependencies and semantic clustering; and Tri-Source Refinement (TSR) to fuse the graph clustering results, CAMs, and input images for boundary refinement. Experiments conducted on the ISIC 2017, ISIC 2018, and PH2 datasets validate the effectiveness of the proposed method.
- 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 behind the paper is relatively clear; it accurately identifies the shortcomings of existing CAM-based weakly supervised methods regarding lesion completeness and boundary quality, and attempts to address these issues simultaneously from three perspectives: global topological modeling, frequency-domain priors, and geometric refinement. 2.The experiments are sufficiently comprehensive, covering three public datasets and comparing the proposed method against various single-stage and two-stage weakly supervised approaches. As indicated in Tables 1 and 2, the method demonstrates consistent improvements in both pseudo-mask quality and final segmentation performance. 3.The ablation studies largely substantiate the effectiveness of each individual module, while Figure 3 visually illustrates the progressive improvement achieved from the baseline model to the full proposed 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.
1.The proposed method comprises multiple components—including frequency-domain modulation, graph clustering, prototype alignment, and tri-source refinement—making the overall framework relatively complex. The paper’s current explanation regarding why these various design elements must coexist could be more in-depth; specifically, the respective contributions and boundaries between the graph clustering module and the subsequent refinement stage warrant a clearer analysis. 2.The current experiments focus primarily on performance metrics themselves; there is relatively insufficient discussion regarding the method’s efficiency, the overhead associated with graph construction, and its scalability when applied to higher-resolution or more complex datasets.
- 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?
The paper features a clearly defined problem, well-motivated methodology, and a relatively comprehensive overall framework. However, as the proposed method comprises multiple components, its overall complexity is relatively high; consequently, the necessity of certain design choices and the relative contributions of individual modules warrant further analysis. Furthermore, the discussion regarding the method’s efficiency, graph construction overhead, and scalability to higher-resolution or more complex datasets is somewhat limited. The paper would be significantly strengthened if the authors could further clarify these issues in their rebuttal or the final version.
- 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 clarifies certain unclear points that was unclear to me.
Author Feedback
We thank the reviewers and meta-reviewer for their constructive comments. We provide the following clarifications and will modify them in the final revision.
Q1[R1] FFM novelty Unlike FFR[25], our FFM does not perform frequency reconstruction or token-level rectification. It uses local spectral priors to modulate backbone features, providing frequency-aware features for subsequent modules.
Q2[R1/MR] TSR and DenseCRF-style refinement TSR is tailored to DGC outputs rather than being a generic post-processing step. It first selects lesion-related clusters via CAM-DGC overlap, then fuses the cluster map with CAM, and finally performs boundary alignment using image geometry. DenseCRF is computationally intensive, and it does not directly perform DGC cluster selection or tri-source fusion with CAM. Therefore, TSR is not trivial or a universal replacement for DenseCRF.
Q3[R1/MR] ViT-based WSSS comparisons There may be a misunderstanding: our comparison is not limited to CNN-based methods. Tables 1 and 2 already include recent ViT-based WSSS methods (ToCo, CoSA, and MoRe). Thus, the reported gains are not merely due to compensating for CNN locality bias but systematic improvement.
Q4[R1/R3/MR] Efficiency and scalability DGC and TSR are training-time supervision-generation components. At inference, only FFM and the trained segmentation branch are used; GNN propagation and TSR refinement are not involved, so DGC and TSR introduce no deployment overhead. During training, DGC constructs graphs on feature maps with an output stride of 16 rather than on original pixels. When the input resolution increases from 224x224 to 512x512, the number of graph nodes only increases from 196 to 1024.Given that our current setting supports a large batch size of 256 on a single consumer-level RTX 4090 GPU, the extra training overhead is manageable.
Q5[R1/MR] Reproducibility and ISIC2018 labels The submitted manuscript states that the code will be publicly released. Labels of ISIC2018 are derived from the “diagnosis_2” field in the official Task 1&2 metadata and merged into four groups similar to the other two datasets. We will release the label CSV and construction script to support reproducibility.
Q6[R1/MR] Cluster number $M$ We clarify that $M$ is not the number of foreground/background classes, but the upper bound of latent structural clusters. We empirically set $M$ = 8 without dataset-specific tuning. This provides moderate over-clustering, as there may be heterogeneous interiors, fuzzy boundaries, occlusions, and background structures. TSR then selects lesion-related clusters via CAM guidance, so redundant clusters do not become final classes.
Q7[R2/MR] Granularity mismatch in FFM We clarify that $F_{mod}$ is not patch-level frequency energy. In Fig.1 (a), $F_{mod}$ denotes the entire modulated deep feature map, combining all patch-wise spectral energy with the whole spatial feature. $F_{mod}$ is then fed into the auxiliary classification head and two other modules (segmentation network and DGC module), constrained by image-level classification, graph-clustering losses, and pseudo-mask supervision; not just matching patch-wise information with one-hot image-level global labels. The auxiliary classification loss regularizes the global semantics of $F_{mod}$ and stabilizes training. We appreciate the suggestion on auxiliary-head ablation or FFM visualization. In the final version, we will replace the current FFM illustration with feature-map visualizations in Fig.1 (a) to make its role clearer.
Q8[R2/R3] Necessity of multiple components coexistence The modules form a progressive design rather than redundant stacking. FFM improves feature reliability with frequency-aware cues; DGC uses the modulated features to recover non-discriminative lesion regions beyond CAMs; and TSR selects lesion-related clusters and aligns pseudo-masks with image geometry. Thus, FFM enhances robustness, DGC improves completeness, and TSR refines boundary consistency.
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 reviewers acknowledge a coherent pipeline with consistent ablation and convincing gains over CNN-based WSSS baselines on ISIC 2017/2018/PH². Two are weakly positive overall, while one is moderately negative, raising two substantive concerns: the absence of a ViT-based WSSS baseline (the field has moved on from ResNet backbones, leaving open whether the GNN clustering is necessary or merely compensating for the CNN’s local inductive bias), and incremental novelty without comparison against standard refinement baselines. A separate methodologically pointed observation from one of the positive reviewers concerns a granularity mismatch in FFM, where patch-level frequency features are supervised only by a global image-level label. The split warrants a rebuttal phase.
For the rebuttal, the most pressing items are comparisons against ViT-based WSSS baselines to address the structural concern raised by the negative reviewer, and a direct response to the FFM granularity-mismatch question — ideally by ablating the auxiliary classification head or by visualizing what the FFM modulation learns spatially. Further items raised by the reviewers — including a DenseCRF refinement baseline, a sensitivity analysis on the number of clusters $M$, computational cost reporting (params, FLOPs, latency at 224 and 512), and clarification of the ISIC 2018 label-extraction protocol — should also be addressed.
- 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 paper proposes a weakly supervised skin lesion segmentation framework that integrates a Frequency-aware Feature Modulation module, a Differntiable Graph Clustering module, and a Tri-Source Refinement module, demonstrating consistent improvements over both single-stage and two-stage WSSS baselines across ISIC 2017/2018 and PH². After rebuttal, R3 recommends acceptance, R2 remains weakly positive, and R1 maintains reject on the grounds of incremental novelty and incomplete benchmarking — most notably the perceived absence of comparisons against ViT-based WSSS methods. I have residual reservations regarding methodological complexity: the combination of three substantial modules could be seen as overcomplicated relative to the underlying technical novelty. However, the rebuttal clarifies that the comparison set in Tables 1 and 2 already includes recent ViT-based WSSS methods (ToCo, CoSA, MoRe), which materially weakens the central premise of the negative recommendation. Given that the most decisive critique does not hold up under scrutiny, and that the empirical contributions across three datasets are consistent and well-supported by ablations, I recommend acceptance on balance.
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.
The paper presents a weakly supervised segmentation framework with consistent gains across multiple skin lesion datasets, supported by ablation studies and comparisons against prior WSSS approaches. Reviewers raised concerns regarding incremental novelty, framework complexity, efficiency, and comparison to stronger transformer-based baselines. The rebuttal clarified several methodological and reproducibility questions, including the inclusion of ViT-based comparisons and the roles of the different framework components. Overall, contribution and experimental evidence are above the acceptance bar.
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.
While the authors’ rebuttal has addressed several of the reviewers’ initial concerns, significant issues remain. In particular, Reviewer 1 continues to express reservations regarding both the novelty of the proposed method and the strength of its empirical validation. Although the paper presents several positive aspects, these outstanding concerns prevent the Area Chair from recommending acceptance. Accordingly, the paper is considered to fall below the acceptance threshold.