List of Papers Browse by Subject Areas Author List
Abstract
Automated skin cancer classification from dermoscopic images remains challenging due to heterogeneous lesion structure, strong intra-class variability, and subtle visual differences between benign and malignant cases. Existing CNN/ViT pipelines typically rely on global or patch-level features and often combine patient metadata via late fusion, which limits spatially grounded multimodal reasoning. We present a novel region-based graph learning framework that explicitly models lesions as graphs of spatially coherent superpixel regions represented as frozen CNN features. To capture fine-grained lesion arrangements, we encode inter-regional geometry as edge attributes and introduce a dedicated metadata context node connected to all regions, providing structured integration of demographic/clinical variables within the same relational space. Node representations are updated using our edge-aware graph transformer followed by attention-driven propagation, and a final graph-level embedding for benign–malignant classification. Experiments on four public benchmarks demonstrate that explicit region-level relational modeling and graph-native multimodal fusion yield consistent gains over the state-of-the-art. Consequently, we establish a new graph-centric perspective in which CNN features are modeled as relational nodes and improved through contextual integration, yielding more expressive and robust classifications.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5100_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/azeemchaudharyg/GeoMeta-GT
Link to the Dataset(s)
HAM10000 Dataset: https://api.isic-archive.com/collections/66/
PAD-UFES-20 Dataset: https://data.mendeley.com/datasets/zr7vgbcyr2/1
ISIC2024 Dataset: https://challenge.isic-archive.com/data/#2024
HIBA Dataset: https://api.isic-archive.com/collections/175/
BibTex
@InProceedings{AzeMuh_GeometryAware_MICCAI2026,
author = { Azeem, Muhammad AND Hussain, Tanveer AND Ahmed, Amr AND Behera, Ardhendu},
title = { { Geometry-Aware Superpixel Graph Transformer with Metadata for Skin Lesion Classification } },
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 a new graph-based approach, GeoMeta-GT, to combine both image features and metadata. The graph is built from nodes corresponding to superpixels (SLIC) in the image, with features given by a frozen CNN, as well as an additional node with metadata features. Then, edges are based on spatial proximity (neighbors) with their features given by geometric attributes, based on distances and orientation. The authors propose a transformer-based message passing strategy that takes into account edge features in the computation of the attention weights, thereby promoting locally adaptive aggregation. The final classification is given by a classification head based on a final feature refinement and graph pooling operations. GeoMeta-GT is applied to skin cancer classification, in a binary normal vs melanoma setup. Results show substantial improvements against other state of the art approaches on multiple benchmark datasets.
- 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 several contributions: 1) a novel geometry-aware message passing algorithm that allows information flow to be conditioned on spatial relationship between nodes, unlike most graph-based approach which often assume that the adjacency matrix itself is representative of the graph geometry; 2) a graph-based approach to fuse superpixel features with metadata. Extensive evaluation on four public datasets validates the proposed approach and shows substantial improvements on the binary classification of dermoscopy images as either benign or malignant.
- 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 validation appears to be done on a single backbone, although the authors do provide an ablation study showing the impact of different backbones in performance. Additionally, most dermoscopy datasets are multiclass, including multiple subtypes of benign and malignant lesions, and part of the real-world clinical challenge differentiating between these subtypes (differential diagnosis). In this work, the evaluation is done only on a binary problem defined as benign vs malignant, therefore simplifying the task.
- 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 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.
(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?
Below is a list of major comments and questions:
- In eq (2), the numerator Qi . Kj + êij suggests that the dot product between Qi and Kj is performed first and then the êij features are added to the result. However, in the text, the authors say that the edge embedding is injected into the key before the dot-product. If the text is correct, putting (Kj + êij) in parenthesis would help clarify.
- In eq (3), why are Dj and eij multiplied by new matrices W4 and W5, instead of the previously used W2 and W3?
- In structural refinement, the authors suggest that the refined embedding Zi is obtained through similarity-based propagation, claiming it allows visually or semantically similar superpixels to exchange information even if they are spatially distant. However, in eq (4), Zi appears to be a combination of neighbors j in N(i), where N(i) is previously defined as the spatial neighbors of node I (based on the Euclidean distance between superpixel centroids), which would contradict the previous claim.
- In the training details, the authors mention a frozen “CNN backbone” without explicitly identifying it. Then, in the ablation studies, they show a comparison between several backbones, but never clarify which backbone ended up being used for the remaining results. Additionally, this opens up questions about whether different backbones would show similar improvements with the proposed GeoMeta-GT.
- In the ablation studies, the authors start with the impact of geometric edge encoding, which is actually shown in Fig 2 (b), while the following section on model variations is shown in Fig 2 (a). I suggest switching the order.
- Fig 2 would also benefit from improved legend, clarifying what the symbols mean - if they actually have any meaning or if color is the only discriminator needed.
- The F1 scores are very similar to the accuracy ones. For clarity and space, consider removing F1 results from the text on the ablation studies without tables or figures. I would also remove it from Fig 2, to simplify my previous comment.
- 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 stand by my original assessment that the paper proposes an interesting approach to fuse image information with metadata using a graph-based strategy and that the proposal is supported by strong experimental validation. The author’s rebuttal addresses most of my concerns and is convincing.
Review #2
- Please describe the contribution of the paper
This paper proposes a Geometry-Aware Superpixel Graph Transformer with Metadata for skin lesion classification, named GeoMeta-GT. GeoMeta-GT designs a geometry-attributed superpixel graph representation and a fusion mechanism that incorporates metadata as graph nodes, and develops an edge-aware graph Transformer to achieve spatially aware aggregation of diagnostically relevant regions. It further introduces a similarity-weighted refinement module to coordinate semantically consistent regional features. Experimental results on multiple public benchmarks show that GeoMeta-GT achieves state-of-the-art performance.
- 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 proposes a dedicated solution for the skin lesion classification task, which is highly meaningful in clinical practice. 2.GeoMeta-GT represents lesion subregions as superpixel nodes, explicitly encodes geometric information on the edges, and integrates metadata as graph nodes. This structured design is more effective than standard feature fusion pipelines. 3.The experimental study is comprehensive, and the ablation experiments provide a detailed analysis of the contribution of each module. 4.The experimental results demonstrate the effectiveness of the proposed method.
- 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 paper only reports some of the main hyperparameters of the proposed model in the Training Details section, while lacking a description of the implementation details for the baselines. This may lead to inconsistencies in the reported performance of different baseline methods. 2.The paper mainly reports statistical metrics such as accuracy, but lacks sufficient clinical analysis. 3.The paper repeatedly claims that the proposed method is robust and generalizable under heterogeneous settings, yet the experiments only include four random train/test splits and do not provide an external validation protocol.
- 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 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 find the main idea interesting: superpixel-based region graphs with geometry-aware attention and metadata-as-node fusion is a plausible and potentially useful contribution for skin lesion analysis. The paper is also reasonably clear and includes detailed experiments. The paper is clearly positioned against patch methods and argues for more spatially grounded multimodal reasoning, which is a sensible direction for dermoscopic analysis.
- 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 author’s rebuttal resolved my doubts. Overall, the experiment was thorough, and I tend to accept it.
Review #3
- Please describe the contribution of the paper
This paper proposes a GeoMeta-GT framework that models lesions as superpixel graphs and integrates patient metadata as a dedicated context node.
- 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 well-organized and easy to understand. GeoMeta-GT achieved SOTA results across ISIC2024, HAM10000, PAD-UFES-20, and HIBA benchmarks. GeoMeta-GT demonstrated superior robustness on challenging, heterogeneous datasets.
- 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.
Superpixel-based graphs for skin lesion diagnosis are well-established in prior researches like SLED [1] and “Seeing the Skin Deeper” [2]. The authors should explicitly clarify their unique technical distinction compared to these existing frameworks.
The proposed designs (Edge-aware Graph Transformer, Structural Refinement and Global Readout) act as general-purpose feature enhancers, lacks a clear link to the unique challenges of Skin Lesion Classification.
The “orientation” attribute versus “distance” remains unclear. An ablation study isolating “distance-only” vs. “full GEE” is required.
The model relies on frozen backbones(e.g.,VGG,EfficientNet) pre-trained on natural images (ImageNet). The authors should address the potential domain gap when applying these frozen features to dermoscopic images and whether domain-specific fine-tuning would significantly alter the performance.
Critical parameters in graph construction(the number of superpixels ($K$) and neighbors ($n$)) lack sensitivity analysis.
The framework’s performance is tethered to the SLIC-based decomposition. The authors should evaluate robustness against different superpixel algorithms.
Minor Weaknesses
The ablation analysis in Fig.2 is difficult to interpret for precise quantitative comparisons. It is recommended to present these results in a tabular format to clearly show the performance delta between model variants.
“Evaluating Backbone Efficiency” section evaluates performance rather than computational efficiency. Notably, the 4% gap between backbones (e.g., on ISIC2024) exceeds the gains from the proposed modules. The authors must ensure that comparisons with SOTA methods used consistent backbones to maintain a fair evaluation.
[1] Zeng G, Peng H, Li A, et al. Unsupervised skin lesion segmentation via structural entropy minimization on multi-scale superpixel graphs[C]//2023 IEEE international conference on data mining (ICDM). IEEE, 2023: 768-777.[2] Jerlshin J G, Pal A R, Datta R, et al. Seeing the Skin Deeper: Interpretable Multi-Task Framework for Skin Lesion Diagnosis using Superpixel Graphs[J]. IEEE Access, 2026.
- 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.
(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?
Please see the major weeknesses.
- 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
Author Feedback
We thank the reviewers and meta-reviewers for their constructive feedback and address the main concerns: novelty, module necessity, graph sensitivity, and evaluation protocol.
Novelty over prior superpixel graphs [Meta, R3]. Superpixel graphs are established in dermoscopy; GeoMeta-GT does not claim novelty from SLIC, but from the reasoning formulation on these regions. SLED is segmentation-oriented, whereas our model performs multimodal benign/malignant classification. Relative to diagnosis-oriented superpixel graph frameworks, GeoMeta-GT: (i) encodes distance/orientation as edge attributes used in transformer attention, so geometry directly conditions message passing rather than appearing only as adjacency; (ii) represents metadata as a context node connected to all regions, allowing patient context to modulate regional evidence via message passing instead of late concatenation; and (iii) CNN for stable descriptors while the graph module for lesion-level relational adaptation. Hence, the contribution is an edge-attributed, metadata-aware graph transformer, not a generic superpixel graph.
Why are geometry and refinement lesion-specific [Meta, R1, R3]? In dermoscopy, malignancy cues are spatial: asymmetry, border irregularity, heterogeneous pigmentation, and repeated suspicious patterns depend on the arrangement of lesion subregions. Geometric edge encoding captures these layouts, and the transformer uses them to weight neighboring evidence. Regarding refinement, as R1 notes, Eq. (4) acts only on the constructed graph neighborhood. It re-weights candidate messages by feature similarity, reinforcing consistent lesion evidence and suppressing noisy superpixels, without creating new long-range edges. This is a clarification of wording/notation, not of the method.
Ablation evidence and orientation [Meta, R3]. R3 concerns about orientation and distance. Distance-only results support the mechanism: full distance+orientation encoding outperforms distance-only on all datasets (ISIC: 98.61 vs 96.17; HAM10000: 98.23 vs 95.23; PAD-UFES-20: 97.17 vs 94.18; HIBA: 95.41 vs 91.13), showing that orientation contributes complementary layout information, particularly for asymmetry and directional border variation.
Graph construction and SLIC sensitivity [Meta, R3]. The concern is dependence on a single SLIC/K/n setting. K controls lesion granularity and n the local context: too small K merges heterogeneous regions, too large K fragments coherent patterns, and too sparse/dense neighborhoods under/over-smooth. Sensitivity over K={15,25,35,50,65,75} and n={4,6,8,10} identifies K=50, n=6 as the best trade-off, with stable mid-range performance. Replacing SLIC with SEEDS (Berg et al., ECCV 2012) or Felzenszwalb (Felzenszwalb et al., IJCV 2004) gives similar results (e.g., 96.17/96.07 on ISIC), showing the transformer is partition-agnostic, though SLIC is preferred for boundary adherence and efficiency.
Backbone, baselines, and robustness [Meta, R1-R3]. The main experiments use ResNet152; the backbone study is a sensitivity analysis, not a protocol change. All ablations share the same preprocessing, splits, input size, optimizer, and backbone, so performance gains reflect graph reasoning, not backbone choice. We use the same binary benign/malignant setting and train/test protocols as the baselines, ensuring comparable evaluation conditions. Results taken directly from prior papers. We concur with R2/R3 that “robust” or “generalizable” does not imply prospective external validation. Our claim is narrower: GeoMeta-GT shows consistent performance across four public dermoscopy datasets and fixed repeated splits. Prospective validation and multiclass differential diagnosis remain future work.
Minor clarifications. Parenthesize edge injection into the key in Eq. (2); in Eq. (3), use separate projections for attention and message update; “Backbone Efficiency” should be “backbone sensitivity”; clarify Fig. 2 to a table and remove redundant F1 plots.
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 proposes a superpixel-based graph Transformer with geometry-aware edge encoding and metadata-as-node fusion for skin lesion classification. The reviews are close to the acceptance boundary, with one Accept, one Weak Accept, and one Weak Reject. The reviewers generally acknowledge the clinical relevance, clear presentation, and strong results across multiple dermoscopy datasets. However, several points require clarification before a final decision, especially the technical distinction from prior superpixel-graph methods, the necessity of the geometry-aware edge encoding and structural refinement modules, and the sensitivity to SLIC decomposition, number of superpixels, and graph-neighbor settings. The authors should also clarify backbone usage, baseline implementation fairness, and whether the claimed robustness is supported beyond random train/test splits. The rebuttal should focus on novelty over existing superpixel graph frameworks, key ablations, graph-construction sensitivity, and fair comparison protocols.
- 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 rebuttal clarifies the technical distinction from prior superpixel-graph methods, provides additional evidence for geometry encoding, and addresses the main concerns about SLIC sensitivity, graph construction, backbone usage, and baseline fairness. Although the contribution remains somewhat incremental and the evaluation is limited to binary classification without prospective external validation, the method is technically coherent and the post-rebuttal reviewer balance supports 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.
Accept
- Please justify your recommendation.
The rebuttal addresses most of reviewers’ concerns.
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.
This paper proposes GeoMeta-GT, a geometry-aware superpixel graph Transformer with metadata fusion for skin lesion classification. The problem is clinically relevant, and the paper presents strong empirical results across multiple public datasets. After considering the reviews and rebuttal, I find that the authors have adequately addressed the reviewers’ main concerns. The overall methodological contribution, strong experimental performance, and comprehensive evaluation are sufficient to support acceptance. Therefore, I recommend Accept.