Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Domain shifts induced by heterogeneous imaging artifacts remain a major obstacle to deploying dermoscopic image classification models in real-world clinical settings. Such artifacts introduce complex domain-specific variations that are difficult to address using domain-invariant representations alone.

In this work, we propose an artifact-aware prompt-based framework for domain generalization in dermoscopic image classification, termed Artifact-aware Prompt Learning for Domain Generalization (APDG). During training, common imaging artifacts are explicitly modeled as multiple source domains, and domain-specific visual prompts are introduced within a Vision Transformer to encode heterogeneous artifact patterns. To mitigate domain bias while preserving useful domain knowledge, we further design a soft-mask-based feature disentanglement module that separates domain-invariant semantic features from domain-specific representations, together with a domain-specific prototype learning mechanism that structurally constrains disentangled features through joint domain supervision and auxiliary class guidance.

At inference time, a test-time prompt fusion strategy is proposed to adaptively aggregate learned domain-specific knowledge without requiring target domain labels or additional training. Experiments on multiple unseen dermoscopic and clinical datasets demonstrate that the proposed method achieves improved robustness and generalization performance under severe artifact-induced domain shifts.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4954_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{JiaXib_Artifactaware_MICCAI2026,
        author = { Jia, Xibin AND Zhang, Wei AND Xu, Shaowu AND Fan, Chao},
        title = { { Artifact-aware Prompt Learning for Domain Generalization } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16881},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes an artifact-aware prompt-based framework (APDG) for domain generalization in dermoscopic image classification that improves the prompt learning using class-aware and domain-specific prompts to address artifact-induced distribution shifts.

  • 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 explicitly models common imaging artifacts as multiple source domains, addressing artifact-induced domain shifts from a principled domain generalization perspective in dermoscopic image analysis. (2) A soft-mask-based feature disentanglement module is designed to separate domain-invariant semantic features from domain-specific representations. The introduction of a prototype-guided mechanism further stabilizes the semantic space and enhances consistency in representations. (3) While prompt learning itself is widely studied, the proposed class-aware and domain-specific prompts provide a novel way to encode heterogeneous artifact patterns, offering a meaningful extension to existing prompt-based methods. (4) A coherent training-to-inference pipeline is established, in which a test-time prompt fusion strategy adaptively aggregates domain-specific knowledge based on prototype similarity, enabling effective transfer to unseen domains without requiring target domain labels.

  • 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 correspondence between modules in the framework figure and the terminology used in the manuscript is not clearly established. For example, the soft-mask-based feature disentanglement (FD) module is labeled as “Mask Generator” in the figure, which may cause confusion. (2) The manuscript does not specify how the learnable prompts are initialized (e. g. , random initialization or based on prior knowledge). Since prompt initialization can affect optimization and final performance, additional details would strengthen the method. (3) Lack of analysis on inference latency. The introduction of test-time prompt fusion (PF) likely incurs additional computational overhead during inference. However, the manuscript provides no quantitative analysis of inference latency (e. g. , runtime or FLOPs). (4) Limited experimental design and evaluation protocol. The experimental setup and presentation are relatively simple. The results are primarily reported via comparison and ablation tables, with AUC as the sole evaluation metric. (5) For the binary melanoma classification dataset, the class distribution is not described. Since medical datasets are often imbalanced, reporting the ratio of positive and negative samples would improve the interpretability of the results.

  • 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?

    The paper demonstrates reasonable novelty, particularly in the design of class-aware, domain-specific prompts and the test-time prompt fusion strategy within a coherent training-to-inference pipeline. However, the overall score is moderated by several limitations in clarity, methodological details, and experimental completeness.

  • Reviewer confidence

    Very confident (4)

  • [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.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper proposes APDG, an artifact-aware prompt learning framework for domain generalization in dermoscopic image classification. The key idea is to treat heterogeneous imaging artifacts (hairs, dark corners, gel bubbles, etc.) as distinct source domains, and to learn domain-specific visual prompts within a Vision Transformer backbone. A soft-mask-based feature disentanglement module separates domain-invariant semantic features from domain-specific representations, regulated by a prototype matrix that jointly supervises class and domain information. At inference time, a test-time prompt fusion strategy aggregates domain-specific prompts via prototype similarity without requiring target domain labels. Experiments are conducted on ISIC 2019 as training source and four independent dermoscopic/clinical datasets as unseen targets.

  • 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 prototype-guided prompt fusion strategy enables adaptation to unseen domains without parameter updates or target-domain labels. This is a pragmatically valuable property for clinical deployment, and the ablation in Table 2 confirms it contributes the largest single performance gain (+1.90% average AUC). 2.APDG outperforms the strongest prior method (EPVT) by 1.34% average AUC across four unseen 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.

    1.The paper treats each artifact type as a separate source domain and relies on artifact annotations from Bissoto et al. (ECCV 2022) to partition ISIC 2019.This is a significant practical constraint that is not adequately discussed. In real clinical settings, obtaining such fine-grained artifact annotations is costly and may not generalize to new acquisition environments. The paper should clarify the annotation cost and discuss how the framework would be applied when such labels are unavailable. 2.Table 2 shows that adding feature disentanglement (FD) on top of domain prompts (DP) actually decreases average AUC from 83.11% to 82.56%. The paper acknowledges this briefly but does not provide a mechanistic explanation. This raises a concern about whether the disentanglement module is well-designed or merely compensated by the subsequent prototype learning step. A more careful analysis, such as visualization of the soft masks or feature space geometry, would strengthen this component’s justification. 3.All experiments involve a single binary classification task (melanoma vs. benign). Dermoscopic diagnosis in practice involves multi-class lesion classification. It is unclear whether the artifact-domain modeling approach would scale to more complex classification settings, and the generality of the framework is therefore limited.

    1. The fusion weights in Eq. (10) are computed by projecting test features onto the prototype matrix B. If the test domain is very different from all source artifact domains, this projection may yield unreliable weights. The paper provides no analysis of fusion weight distributions, prototype discriminability, or failure cases where the fusion strategy underperforms direct domain prompt selection.
  • 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?

    APDG addresses a well-motivated problem and the test-time prompt fusion strategy is a practically valuable contribution that yields the largest performance gain in the ablation. The overall results are competitive. However, several issues limit the paper in its current form. The requirement for artifact domain annotations during training is a significant practical constraint that is insufficiently addressed. The feature disentanglement module demonstrably hurts performance when used alone, raising questions about its design. The evaluation is restricted to binary classification on a single training source.

  • Reviewer confidence

    Very confident (4)

  • [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.

    After considering the rebuttal, I raise my score to 4 (Weak Accept). While the artifact annotation dependency and binary classification scope remain limitations, these do not fundamentally invalidate the core technical contribution. The overall results are competitive across four unseen datasets.



Review #3

  • Please describe the contribution of the paper

    This paper presents a novel artifact-aware approach to dermoscopic image domain generalization by explicitly modeling heterogeneous imaging artifacts as multiple source domains, thereby capturing real-world variability more effectively. Building on this formulation, the authors propose a unified prompt-based framework that integrates domain-specific prompts, feature disentanglement, and prototype-guided learning to separate lesion-relevant semantics from artifact-induced variations and preserve clinically meaningful representations across domains. Furthermore, the authors introduce an adaptive test-time prompt fusion strategy that leverages learned domain knowledge without requiring target-domain labels, enabling robust generalization to unseen distributions.

  • 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.Novel and clinically grounded problem formulation: The paper introduces an artifact-aware domain generalization setting by modeling heterogeneous dermoscopic artifacts as multiple source domains, which more realistically reflects clinical imaging variability and addresses an important gap in existing DG formulations. 2.Methodological advancement over prior prompt-based DG approaches: The proposed framework improves upon existing prompt-based methods by explicitly enforcing feature disentanglement between artifact-induced variations and lesion-relevant semantics, leading to more robust representations. 3.Test-time generalization: The adaptive prompt fusion strategy enables label-free test-time adaptation to unseen domains, which is commendable, as target data labels are not always available during testing, especially when comparing across multiple datasets. 4.The proposed model demonstrates consistent performance gains under severe domain shifts across different 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.

    1.Limited evaluation across centers and real-world settings: The experimental validation does not include a true multi-centre evaluation or sufficiently diverse acquisition settings. This raises concerns about how well the proposed method generalizes to broader clinical environments with varying devices, protocols, and patient populations.

    2.Narrow evaluation metrics: The performance assessment relies primarily on AUC, which, while informative, does not provide a complete picture of clinical utility. Additional metrics such as sensitivity, specificity, F1-score, calibration, and decision-curve analysis would strengthen the evaluation.

    3.Incomplete comparison with domain generalization baselines: The study does not include a sufficiently comprehensive set of state-of-the-art DG methods beyond prompt-based approaches. In particular, several well-established DG techniques (e.g., meta-learning, data augmentation-based methods, or invariant learning methods) are not considered, which limits the strength of the comparative analysis.

  • 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?

    The methodology is novel, a well-motivated artifact-aware domain generalization framework that meaningfully advances prompt-based learning by explicitly disentangling lesion semantics from domain-induced variations, leading to improved robustness under clinically relevant domain shifts. The authors provided sufficient information for reproducibility.

  • 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

Reviewer #1 Q1 Terminology consistency The “Mask Generator” in Fig. 1 is a part of the FD module. We will unify the terminology in the text and figures. Q2 Prompt initialization Domain-specific prompts are randomly initialized. We will clarify this in the experimental setup. Q3 Inference overhead We will add the requested quantitative FLOPs/runtime analysis in the revision. Unlike other methods, our prompt fusion involves no parameter updates at test time. Benefiting from efficient matrix operations and the limited number of prompts and prototypes, generating the fused prompt introduces only a very small additional computational overhead in terms of FLOPs and runtime. We will include the results in the final version. Q4 Evaluation metrics ROC-AUC is widely adopted in melanoma classification and domain generalization settings, particularly under class imbalance. Therefore, following prior work, we use ROC-AUC as the primary evaluation metric. We agree that metrics such as sensitivity analysis could provide complementary clinical perspectives. We will add these additional evaluation metrics in the final version. Q5 Class distribution We will add the melanoma/benign class distribution in Section 4.1 (Dataset description) of the revised manuscript. Reviewer #2 Q1 Artifact annotations and practical applicability As discussed in the Introduction, the overfitting to artifacts leads to reduced generalization performance. We use the ECCV 2022 artifact annotations to partition ISIC2019, which helps address this problem. In real clinical settings, APDG can be extended to multi-center DG using clinical centers as source domains, requiring only simple domain identifiers (e.g., hospital indices). We will further discuss this point in the revised manuscript. Q2 Effect of FD FD and PL are designed as a unified, synergistic mechanism. FD performs feature disentanglement via soft masks, while PL provides auxiliary supervision signals (L_d, L_c, L_mix) during the disentanglement process. Therefore, applying FD alone without PL leads to a performance drop. We will further clarify this mechanism in the ablation study section. Q3 Binary classification setting Our architecture is mathematically agnostic to the number of classes, as the class mapping matrix M_cls seamlessly can naturally extend to multi-class settings without structural modifications. We will explore this in future multi-class extensions. Q4 Fusion on Extreme Target Domains Our prompt fusion is designed with an inherent safeguard. Because the fusion weights are computed via a Softmax operation, if a test sample encounters highly unfamiliar artifacts, its features tend to yield uniformly low similarity scores across all learned source prototypes. The Softmax operation helps flatten these scores into a near-uniform distribution. This allows the model to fall back to a generalized, domain-agnostic ensemble representation. Reviewer #4 Q1 Multi-centre and real-world evaluation We agree that broader multi-centre evaluation is important for assessing real-world robustness. In this work, we evaluated APDG on four independent unseen datasets collected under different acquisition conditions and imaging protocols, aiming to simulate clinically relevant domain shifts. We will explore larger-scale multi-centre studies in future work. Q2 Evaluation metrics ROC-AUC is widely adopted in melanoma classification and domain generalization settings, particularly under class imbalance. Therefore, following prior work, we use ROC-AUC as the primary evaluation metric. We agree that metrics such as sensitivity analysis could provide complementary clinical perspectives. We will add these additional evaluation metrics in the final version. Q3 DG baseline comparison We compare with representative methods closely related to our approach, including optimization-based, regularization-based, and prompt-based DG methods. We will include additional domain generalization methods in the revised manuscript.



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 paper proposes an artifact-aware prompt learning framework for domain generalization in dermoscopic image classification. Reviewers find the problem formulation well-motivated and appreciate the label-free test-time adaptation. However, they share concerns about the reliance on costly artifact annotations, unclear methodological details, and questionable behavior of the feature disentanglement module. The evaluation is also viewed as limited, restricted to binary classification with AUC as the sole metric and missing broader DG baselines. Given these mixed concerns, the paper moves to 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 rebuttal adequately addressed reviewers’ concerns, leading one reviewer to raise their score. The recommendation is therefore Accept.



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 is recommended for acceptance based on its novel, clinically grounded formulation and convincing performance in mitigating artifact-induced distribution shifts for dermoscopic image classification. Reviewers strongly credited the prototype-guided prompt fusion strategy, which elegantly enables label-free test-time adaptation to unseen domains without requiring target labels or parameter updates



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.

    Following the rebuttal, the reviewers reached a consensus to accept the paper, agreeing that its introduction of artifact-specific prompt design is well-motivated for a domain generalization method. I therefore recommend acceptance.



back to top