SkinGPT-R1: Adapter-Based Dermatologic VLM

• SkinGPT-R1 is an adapter-based vision–language model that combines a frozen Vision-R1-7B backbone with two efficient, trainable adapters to enable dermatologist-grade reasoning.
• It leverages a large, curated DermCoT corpus and dual-distillation techniques for state-of-the-art clinical performance in both reasoning and classification tasks.
• The framework ensures explicit chain-of-thought outputs, enhancing interpretability and enabling rapid, deployable diagnostic solutions in clinical settings.

SkinGPT-R1 is an adapter-based vision–LLM (VLM) targeting dermatologic diagnostic reasoning with explicit, verifiable chain-of-thought (CoT) capabilities. Developed on the Vision-R1-7B backbone, it leverages dermatologist-verified narrative supervision and efficient, frozen-weight adaptation via two trainable adapters. The framework introduces DermCoT—a large, curated corpus of standardized dermatologic CoT narratives—and integrates both dual distillation (visual and CoT) and multi-dimensional clinical evaluation. Comprehensive benchmarking demonstrates state-of-the-art performance across clinical reasoning and classification tasks while maintaining high efficiency and deployment viability (Shen et al., 19 Nov 2025, Liu et al., 18 Nov 2025).

1. Architectural Framework and Adapter Mechanisms

SkinGPT-R1 employs the Vision-R1-7B model as its backbone, retaining all core parameters in a frozen state and introducing two parameter-efficient adapters along the vision-to-language pipeline:

• Visual Alignment Head (“adapter”): A residual network that projects frozen image patch features into a dermatologist-trained teacher embedding space, $\mathbf{t} \in \mathbb{R}^{D_t}$.
• Low-rank Language Bias Adapter: Converts the compact image summary into the decoder hidden space, providing an additive bias $\mathbf{b}$ on the vocabulary logits at supervised positions.

Both adapters are identity-initialized to ensure training stability and to allow for rapid convergence towards dermatology-specialized reasoning without disturbing base VLM capabilities.

Dual-Distillation Objective

SkinGPT-R1’s learning process combines two modalities:

• Visual Distillation: Minimizes mean-squared error between student ($\mathbf{z}$) and fixed teacher embeddings ($\mathbf{t}$):

$$\mathcal{L}_\text{vis} = \frac{1}{B}\sum_{b=1}^B \|\mathbf{z}^{(b)} - \mathbf{t}^{(b)}\|_2^2$$

• CoT Supervision: Next-token prediction cross-entropy with dermatologist-structured CoT target $y^*$:

$$\mathcal{L}_\text{cot} = -\frac{1}{|T|}\sum_{t\in T} \log p_\theta (y^*_t \mid y^*_{<t},x)$$

• Total Loss: Curriculum-scheduled linear combination:

$$\mathcal{L}_\text{total} = \lambda_\text{vis}(s) \mathcal{L}_\text{vis} + \lambda_\text{cot}(s) \mathcal{L}_\text{cot}$$

Here, $s$ denotes normalized training step with cosine-scheduled weights $\lambda_\text{vis}$ and $\lambda_\text{cot}$ (Shen et al., 19 Nov 2025).

2. Dermatology-Specific Training Data: DermCoT Corpus

The DermCoT corpus supports the domain-specific reasoning capability of SkinGPT-R1 through large-scale, high-quality, and balanced dermatologic narratives:

• Certified Test Set: 3,000 image–CoT pairs manually reviewed and corrected by board-certified dermatologists.
• Filtered Training Candidates: 15,000 auto-drafted image–CoT pairs, evaluated by DermEval. The top 10,000 with mean score $\geq 4.5/5$ are retained, enforcing balancing across diagnosis classes and anatomical sites.

CoT Narrative Structure

Each narrative is constructed in three standardized layers:

1. Observation-Only Caption: Describes anatomic site, morphology, distribution, color, and surface changes; diagnostic leakage is explicitly avoided.
2. Label-Aware Hierarchical Draft: Accumulates discriminative evidence before revealing the ground-truth label.
3. Template Normalization: Outputs a strictly-layered CoT (visual findings $\to$ reasoning $\to$ diagnostic conclusion) with terminology control.

This compositional constraint ensures both interpretability and downstream evaluability (Shen et al., 19 Nov 2025).

3. Clinical Evaluation: DermEval and DermBench

Evaluation is centered on both data curation and output quality, mapped to clinical priorities via multi-dimensional scoring.

DermEval

• Function: Automated curation system based on a LLaVA-architecture. Trained in two stages: supervised six-score output formatting and REINFORCE-based alignment to physician labels (reward $= -$MSE).
• Application: Filters draft narratives, selecting the subset for training with maximal clinical alignment.

DermBench

• Definition: Model evaluation benchmark using the 3,000 certified DermCoT cases.
• Method: Each candidate model's output is compared (via VLM “comparator”) to gold-standard reference narratives using a fixed-prompt and temperature setup.
• Scoring Criteria (scale: 1–5):

1. Accuracy (diagnosis/key findings match reference)
2. Safety (absence of harmful advice)
3. Medical Groundedness (domain knowledge support)
4. Clinical Coverage (findings, differentials, plan)
5. Reasoning Coherence (logical stepwise structure)
6. Description Precision (clarity, term accuracy)

• Metric: System-level scores are averaged across the six criteria and 3,000 cases (Shen et al., 19 Nov 2025).

4. Empirical Results and Ablation Studies

Clinical Reasoning Performance

On DermBench, SkinGPT-R1 achieves an average of 4.031/5 across all six dimensions, outperforming its frozen Vision-R1 backbone (2.865/5) by approximately 41% in relative gain. Notable category scores include Clinical Coverage (4.403), Description Precision (4.637), and Safety (4.187). SkinGPT-R1 leads or matches in nearly all clinical reasoning dimensions (Shen et al., 19 Nov 2025).

Zero-Shot Dermatology Classification

Stable accuracy gains are observed on diverse skin disease benchmarks:

Benchmark	Vision-R1	SkinGPT-R1
Derm7pt	27.3%	32.9%
PAD-UFES-20	31.7%	37.6%
Skin Lesion (39 cls)	7.0%	8.6%

This demonstrates broad applicability and generalization across disease class taxonomies (Shen et al., 19 Nov 2025).

Ablation Analyses

Sequential addition of narrative CoT and visual distillation shows cumulative gains:

Configuration	Avg Score DermBench	Zero-Shot Acc Boost
Vision-R1 (no CoT, no distill)	2.865	—
+ CoT only	3.134	+2–3 pts
+ CoT + Visual Distillation (SkinGPT-R1)	3.476	+3–4 pts

CoT supervision yields the largest jump, with visual distillation providing consistent additional improvements.

5. Efficiency, Interpretability, and Deployment

SkinGPT-R1 is parameter-efficient and deployable in real-world clinical environments:

• Only the two adapters and a scalar gate are trained; the main backbone remains frozen. This enables rapid adaptation and preserves inference speed and memory profile identical to Vision-R1.
• Identity initialization and mixed-precision training permit stable learning on standard GPU clusters.
• Explicit, layered CoT output separates visual observation, reasoning, and diagnosis, facilitating clinical auditability and reducing the “black-box” profile typical in VLMs (Shen et al., 19 Nov 2025).

6. Context within Dermatology Reasoning Models

While SkinGPT-R1 emphasizes adapter-based, verifiable CoT distillation and lean adaptation, alternative approaches such as SkinR1 (aka SkinGPT-R1) adopt a unified end-to-end paradigm with supervised fine-tuning (SFT) on textbook-derived reasoning trajectories, followed by reinforcement learning (RL) with group-relative policy optimization (GRPO) (Liu et al., 18 Nov 2025). SFT imparts expert-level, hierarchy-aware reasoning; GRPO propagates this skillset to large, sparse datasets, further enhancing accuracy and robustness. Empirically, SkinR1 demonstrates in-distribution accuracy of 0.6385 and out-of-distribution accuracy of 0.7171, surpassing standard VLMs and competitive baselines in both clinical trustworthiness and structural compliance (Liu et al., 18 Nov 2025).

7. Significance and Implications

SkinGPT-R1 establishes a new standard for transparent, adaptable, and auditably accurate dermatologic reasoning in VLMs. Its coupling of dermatologist-certified chain-of-thought narratives with parameter-efficient visual distillation enables both strong clinical performance and real-world deployability. The framework's explicit reasoning pathways, multi-dimensional evaluation, and integration with automated curation further address historical limitations of reasoning opacity, data heterogeneity, and transferability in clinical vision–language systems (Shen et al., 19 Nov 2025, Liu et al., 18 Nov 2025).