LLM-Based Visual Explanation Evaluation Framework for Assessing the Explainability of Facial Skin Disease Classification Models

Gyuyeon Na

REVIEW 4 major objections 4 minor 2 references

LLM-Based Visual Explanation Evaluation Framework for Assessing the Explainability of Facial Skin Disease Classification Models

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that an LLM-as-a-Judge framework, steered by progressively richer prompts, can quantitatively assess whether Grad-CAM heatmaps from facial skin disease classifiers point at clinically relevant lesion regions, and that doing

desk verdict A candid pilot with a plausible classification study but an unsupported central claim: the LLM localization scores are likely prompt compliance, not spatial perception, because the clinical priors are injected into the prompt itself. read the letter →

arxiv 2606.16794 v2 pith:OZ2ILV3U submitted 2026-06-15 cs.CV

Gyuyeon Na This is my paper · ORCID

classification cs.CV

keywords LLM-as-a-JudgeexplainableAIGrad-CAMfacialskindiseasemedicalimagingpromptengineeringattentionlocalizationdataaugmentation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large language models can act as quantitative judges of visual explanations in medical imaging, according to this pilot study. The paper proposes an LLM-as-a-Judge framework that scores Grad-CAM heatmaps from facial skin disease classifiers on two criteria—how well the heatmap localizes the true lesion and how trustworthy the explanation is—using a five-point scale. Its key move is a progressive prompt design: starting from a basic request, it adds scoring rubrics, clinical knowledge about where facial atopic dermatitis appears, penalty rules for irrelevant activations, and a structured JSON output. The authors report that scores rise and become more consistent across three LLMs as the prompt gets richer, and that final scores closely match ratings from three non-expert human evaluators. If the claim holds, researchers gain a repeatable, low-cost way to audit whether a medical AI model is actually looking at the disease.

What carries the argument

The load-bearing mechanism is the progressive prompt engineering strategy (P1–P5), applied inside an LLM-as-a-Judge evaluation. Each stage adds a distinct constraint to the prompt: a dermatologist role assignment, a five-point rubric for Localization and Trustworthiness, clinical priors stating that facial atopic dermatitis typically appears around the eyes, forehead, glabella, and cheeks, penalty rules for activations in background, accessories, clothing, or non-lesion regions, and finally a structured JSON output. These staged additions are what the authors credit for increasing the consistency, clinical grounding, and reproducibility of the LLM scores. The framework's two evaluation crite

What would settle it

Take a Grad-CAM heatmap from a model trained on a different disease (or an intentionally misaligned heatmap that highlights background), present it to the same LLMs with the same prompts, and check whether Localization scores drop as expected; if scores stay high or move only in response to the clinical text, the framework is reading its priors rather than the image. A second falsifier is to compute pixel-level overlap (e.g., Dice or IoU between the heatmap threshold and a manual lesion segmentation) and compare it with the LLM scores across a set of images—no correlation would mean the LLM is

Watch

Extended reading notes

Core claim

The paper's central discovery, stated on its own terms, is that a domain-adapted LLM-as-a-Judge prompt can produce consistent quantitative scores for the quality of Grad-CAM explanations in facial skin disease classification. The authors show that adding progressively richer instructions—rubric definitions, clinical priors about lesion locations, penalties for background activations, and structured output—increases the Localization and Trustworthiness ratings assigned by GPT-5.5, Gemini 3.5 Flash, and Claude Sonnet 4.6, and makes the ratings more stable across models. They further report that the most refined prompt yields LLM ratings that closely track the average of three non-expert pilot

Load-bearing premise

The framework's validity rests on the assumption that the LLMs are genuinely perceiving the spatial overlap between the Grad-CAM heatmap and the true lesion region, rather than echoing the clinical priors (eyes, forehead, glabella, cheeks) that the prompt itself injects in stage P3; if the latter, the reported Localization scores measure prompt compliance, not explanation quality.

Editorial extensions

If this is right

If LLM judges prove reliable, medical imaging teams can audit Grad-CAM explanations at scale without requiring a dermatologist to review every heatmap.
The finding that augmentation strategy changes where models attend implies that accuracy alone is an insufficient model-selection criterion; an explainability check should accompany augmentation choice.
Because the best augmentation differs by architecture (mixed for EfficientNet-B0, geometric for MobileNetV3, color for ResNet18), default or transferred augmentation recipes may silently degrade both classification and explanation quality.
The fact that injecting clinical priors raises localization scores means prompt content is a variable that any future LLM-based medical evaluation must report and control.
The observed similarity between LLM scores and non-expert human scores suggests LLMs could act as a screening filter that flags low-quality explanations before expert review.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A plausible alternative reading of the prompt-refinement effect is that LLMs are largely parroting the clinical priors in the prompt rather than measuring spatial overlap; a decisive test would be to feed the same heatmap with deliberately shifted clinical priors and check whether localization scores follow the text or stay anchored to the heatmap.
The rising scores from the basic to the structured prompt could reflect prompt compliance rather than improved evaluation accuracy; whether 'higher' means 'better' or just 'more aligned with the prompt's expectations' is unresolved without ground truth.
Because the study uses one hand-picked image and one model·augmentation combination, the framework's generality is untested; a natural extension is to run the same prompt stack on a diverse image set with manual lesion segmentations and correlate LLM localization scores with pixel-level metrics such as Dice or IoU.
The findings suggest a practical extension: a 'heatmap sanity check' service that takes any Grad-CAM output and returns a localization score could be built on this prompt stack, but its calibration against expert judgment would be the make-or-break question.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

What you should know: this is a single-author pilot that combines a straightforward augmentation/classification study with an LLM-as-judge evaluation of Grad-CAM explanations. The classification half is fine. The explainability half does not support the paper's headline contribution.

The new thing here is the specific object: using multimodal LLMs to score Grad-CAM localization and trustworthiness for facial skin disease, with a progressively engineered prompt set (P1–P5). That specific combination doesn't appear in the cited literature, and the paper does a fair job of describing a method that could become a cheap screening tool. It also honestly concedes the big limitations: one image, three non-expert raters, no expert ground truth, manual web-interface collection. Those concessions kept me from dismissing it outright.

The classification experiments are credible. Standard backbones, four augmentation settings, repeated runs with mean±std, and architecture-dependent conclusions that stay close to the data. The results are internally coherent. The class-wise recall analysis is reasonable.

But the central claim—that these LLM scores quantitatively assess whether a Grad-CAM heatmap is grounded in the actual lesion—is not demonstrated. The stress-test note is right. P3 tells the LLM that atopic dermatitis appears around the eyes, forehead, glabella, and cheeks. P4 tells it to penalize background activations. The scores rise monotonically from P1 to P5, which is exactly what a prompt-compliant system would do even if it were merely reading the text and ignoring the heatmap geometry. The human pilot doesn't break the confound because the raters got the same rubric and presumably the same clinical prior. There is no negative control: no shuffled overlays, no deliberately mislocalized heatmaps, no lesion-free image. Without that, the Discussion claim that clinical prompts produced "more clinically grounded" assessments is self-confirming.

The paper has a few other soft spots. The validation set appears to serve as the test set, the number of repeated runs is not stated, and there is no code or data. The dataset is not identifiable from the citation. Also, the abstract describes a different attention generation algorithm (color saliency plus Gaussian smoothing) than the Grad-CAM used in the body. That's a mechanical inconsistency that must be fixed.

Is the paper worth referee time? Yes, in the sense that the question is timely and the author is being honest about limits. But it needs heavy revision: multiple images, expert ground truth, negative controls, a real test set, and resolution of the abstract mismatch. As is, the central claim is a hypothesis, not a result. I wouldn't cite it in its current form, though it could be a useful reading-group case study on confounds in LLM evaluation.

Recommendation: send to peer review with the expectation of major revision or rejection, but do not desk-reject. The idea has merit and the limitations are stated; the missing controls are fixable.

Referee Report

4 major / 4 minor

Summary. The paper proposes an LLM-as-a-Judge framework for quantitatively assessing Grad-CAM visual explanations in facial skin disease classification. Three CNN architectures (EfficientNet-B0, MobileNetV3, ResNet18) are trained under four augmentation conditions, and Grad-CAM heatmaps from a single representative atopic dermatitis image are evaluated by GPT-5.5, Gemini 3.5 Flash, and Claude Sonnet 4.6, as well as by three non-expert human raters. A five-stage progressive prompt strategy (P1–P5) is introduced, adding evaluation rubrics, clinical knowledge, penalty rules, and structured output. The paper reports that augmentation effectiveness depends on model architecture, that models under their best augmentation focus more on lesion regions, and that progressively richer prompts increase LLM scores, which is interpreted as improved consistency and clinical grounding.

Significance. If the central claim were established, the framework would offer a scalable, quantitative complement to qualitative inspection of XAI outputs in medical imaging. The classification comparison across architectures and augmentations is a useful empirical contribution, and the authors are transparent about the pilot nature of the study. However, the core evaluation claim is not yet supported: no experiment distinguishes genuine spatial perception from compliance with the clinical priors injected in the prompt, no expert-derived ground truth anchors the scores, and the entire LLM evaluation rests on a single image. The paper is best read as a proof-of-concept proposal that requires substantial additional validation.

major comments (4)

[§3.2.5, §4.2, Table 4] The central claim that clinical knowledge improves clinically grounded localization assessment is confounded by prompt compliance. P3 injects the expected lesion locations (eyes, forehead, glabella, cheeks) and P4 explicitly penalizes activations in backgrounds and accessories. The observed monotonic increase in Localization (e.g., GPT-5.5: 3.85 at P1 to 4.70 at P5) is exactly what a system with no visual grounding would produce if it simply reflected the prompt's priors. The paper provides no control condition—e.g., a deliberately wrong or shifted heatmap, a random heatmap, or a mismatched image-heatmap pair—to test whether the LLMs can detect spatial correspondence rather than echoing the prompt. This is load-bearing for the framework's validity.
[§3.2.3, §4.2, Table 3] The proposed framework is evaluated on a single representative atopic dermatitis image, generated from one model·augmentation combination (MobileNetV3 with Augmentation1). No other disease category, image, model, or augmentation setting is included in the LLM evaluation, despite the classification experiments covering three models and four augmentation strategies. The conclusion that the framework is generally applicable to Grad-CAM assessment in facial skin disease classification is therefore unsupported by the evidence. The limitation is acknowledged in §6, but the framing in §5 still presents the framework as the paper's primary contribution.
[§3.2.6, §4.2, Table 3, §6] The human pilot evaluation does not break the circularity. The three non-expert evaluators were given the same rubric and the same clinical prior information as the LLMs, so agreement between LLMs and these raters cannot validate clinical grounding. Section 6 explicitly states that expert-derived ground truth was not available and that the absolute accuracy of LLM evaluations could not be validated. To support the framework, the authors need an external anchor—for example, dermatologist annotations, or an objective spatial overlap metric (IoU or Dice) between thresholded Grad-CAM regions and lesion segmentations. Without such an anchor, the reported similarity of scores in Table 3 is not evidence of validity.
[§3.2.2, Table 1] The classification results are reported as mean ± std, but the number of repeated runs is never stated. In addition, the same validation set appears to be used both for early stopping and for selecting the best augmentation condition, with no separate test set reported. This makes the reported 'best augmentation' claims and the subsequent explainability analysis potentially optimistic or overfit. The run count and a clear train/validation/test split should be specified.

minor comments (4)

[Table 3] The reported standard deviations for LLM scores are unclear: with temperature=0, variability could arise across prompt stages, repeated queries, or something else. The table caption should state what the ± values represent.
[Fig. 1] The Grad-CAM visualization would benefit from a color scale and a thresholded contour overlay to make the claimed spatial overlap between activation and lesion visibly quantifiable.
[References] Several references are incomplete (e.g., Wei et al. 2023 has no arXiv identifier; Yang et al. 2023 and Shen et al. 2022 lack volume/page details). The AI-Hub dataset should also be formally cited or described with an accession or version.
[§4.2] The phrase 'consistent and stable' is used to describe the effect of prompt engineering, but Table 4 shows single numeric scores per prompt stage, not repeated-measure variability. Please clarify the basis for the consistency/stability claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the framework is empirical and its limitations are explicitly acknowledged.

full rationale

The paper does not derive a result from its inputs by construction. It trains three CNN baselines, generates Grad-CAM heatmaps, and asks LLMs to rate them under five prompt variants. The monotonic P1-to-P5 score increase (Table 4) is framed as evidence that rubrics, clinical priors, penalty rules, and structured output improve evaluation consistency; this is a prompt-ablation observation, not a prediction made from a fitted parameter. The agreement with the pilot human raters (§4.2) is a comparison against raters who used the identical rubric, but that weakens external validity rather than making the conclusion equivalent to the input: both sets of raters could disagree, and no formal constraint forces the reported agreement. The paper explicitly concedes that no expert-derived ground truth was available and that the study is a pilot on a single image (§6), so the authors do not claim an independently validated measure of clinical localization. No self-citation, uniqueness theorem, or imported ansatz is load-bearing. The strongest concerns raised by a skeptical reader—that the LLMs may be parroting clinical priors injected in P3/P4, and that the human raters share the same rubric—are real limitations of evidence, but they are concerns about external validity and missing control conditions, not a demonstration that any output reduces by definition to the inputs.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

The framework contributes no fitted equations: its quantitative output is entirely mediated by hand-authored rubric anchors, clinical priors, and penalty rules inserted into LLM prompts, and by hand-chosen training/augmentation hyperparameters. The load-bearing axioms are (1) Grad-CAM heatmaps reflect the model's true decision basis, (2) LLMs can visually arbitrate heatmap-to-lesion correspondence, (3) the P3 clinical priors are the correct ground truth for the test image, and (4) same-rubric non-expert ratings are a meaningful cross-check. The paper provides no external anchor (expert ground truth) to break the loop between the injected standard and the measured score.

free parameters (7)

learning_rate = 0.0001
Hand-chosen in §3.2.2; standard training hyperparameter, not fitted to data.
batch_size = 32
Hand-chosen in §3.2.2.
max_epochs / early-stopping patience = 7 / 2
Hand-chosen in §3.2.2; defines the stopping rule used for all models.
geometric augmentation parameters = RandomResizedCrop(0.85, 1.0), RandomHorizontalFlip(p=0.5), RandomRotation(10°)
Hand-chosen in §3.2.1; defines the Augmentation1 condition that drives the MobileNetV3 conclusions.
color augmentation parameters = ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.02)
Hand-chosen in §3.2.1; defines the Augmentation2 condition.
LLM generation temperature / top_p = 0 / 1.0
Set in §3.2.4 to force deterministic-looking outputs; the consistency claims depend on these settings.
Likert rubric anchors = 5-level anchors for Localization and Trustworthiness
Authored by the authors (Fig. 2, §3.2.5); the entire evaluation output is mediated by these hand-defined anchors.

assumptions (7)

domain assumption Grad-CAM heatmaps faithfully represent the model's decision-relevant regions.
Invoked in §3.2.3 and §4.1, where heatmap focus is equated with 'the model relied primarily on disease-related skin characteristics' — an assumption debated in XAI literature and not tested here.
domain assumption Multimodal LLMs can accurately perceive and judge spatial overlap between a heatmap and the true lesion from static images.
The load-bearing premise of the framework (§3.2.4, §4.2); no independent test of visual correspondence is provided.
domain assumption The P3 clinical priors (atopic dermatitis appears around eyes, forehead, glabella, cheeks) are correct and complete for the test image.
Injected in §3.2.5 (P3); if wrong for this image, the localization scores are anchored to a false standard.
domain assumption ImageNet-pretrained initialization transfers to facial skin disease classification.
Standard transfer-learning assumption used in §3.2.2.
domain assumption The AI-Hub synthetic facial skin disease images are representative of clinical imaging conditions.
Stated in §3.1; no external validation or dataset identifier is provided.
domain assumption Non-expert human ratings using the authors' rubric are a meaningful cross-check for LLM evaluation.
Used in §3.2.6/§4.2 as the comparison benchmark; same rubric as the LLMs, so agreement is partly shared-criteria agreement.
domain assumption temperature=0 with top_p=1.0 yields deterministic, reproducible LLM outputs.
Stated in §3.2.4; in practice web-interface/model versions are not bit-reproducible, and the paper's own reproducibility section notes manual collection.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LLM-Based Visual Explanation Evaluation Framework for Assessing the Explainability of Facial Skin Disease Classification Models." pith.science (2026). https://pith.science/paper/OZ2ILV3U

@misc{pith2026260616794,
  author       = {Pith},
  title        = {Pith review of: LLM-Based Visual Explanation Evaluation Framework for Assessing the Explainability of Facial Skin Disease Classification Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZ2ILV3U}},
  note         = {Machine review of arXiv:2606.16794}
}

read the original abstract

This study proposes a domain-specific LLM-based Visual Explanation Evaluation Framework for assessing visual attention explanations in facial skin disease diagnosis. While previous studies have primarily focused on improving classification performance, relatively few studies have systematically examined whether visual explanations are grounded in clinically relevant lesion regions. In this study, an image-driven visual attention generation algorithm was developed to produce lesion-focused attention maps from facial skin disease images. Unlike conventional gradient-based explainability methods, the proposed approach combines color saliency, facial spatial priors, Gaussian smoothing, and attention overlay visualization to generate clinically interpretable attention maps. Furthermore, an LLM-as-a-Judge evaluation framework was designed using GPT-5.5, Gemini 3.5 Flash, and Claude Sonnet 4.6 to assess the generated visual explanations from the perspectives of lesion localization and explanation trustworthiness.

Figures

Figures reproduced from arXiv: 2606.16794 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 linked inside Pith

[2022]

Computers in Biology and Medicine, 149:105939

An Improved Transformer Network for Skin Cancer Classification. Computers in Biology and Medicine, 149:105939
[2025]

Journal of Imaging

Skin Lesion Classification Through Test Time Augmentation and Explainable Artificial Intelligence. Journal of Imaging. Yang, B. et al. 2023. A Deep Learning-Based Diagnostic Model for Skin Pigmentary Diseases in an IoT Smart Healthcare Environment. In Proceedings of the International Symposium on Artificial Intelligence and Medical Sciences. Shen, S. et a...

arXiv 2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [2022]

Computers in Biology and Medicine, 149:105939

An Improved Transformer Network for Skin Cancer Classification. Computers in Biology and Medicine, 149:105939

[2] [2025]

Journal of Imaging

Skin Lesion Classification Through Test Time Augmentation and Explainable Artificial Intelligence. Journal of Imaging. Yang, B. et al. 2023. A Deep Learning-Based Diagnostic Model for Skin Pigmentary Diseases in an IoT Smart Healthcare Environment. In Proceedings of the International Symposium on Artificial Intelligence and Medical Sciences. Shen, S. et a...

arXiv 2023