DermoBench - Morphology: leaderboard
Dermatology image understanding from DermoBench (DermoGPT paper, Ru et al., 2026): the Morphology axis averages two open-ended lesion-description tasks, scored 0-100 by a Gemini-2.5-Pro judge against dermatologist-revised references on a 900-image core set, and two attribute multiple-choice tasks built from the Derm7pt dermoscopy checklist and SkinCon clinical concepts. The score is the paper's mean of the four, judge scores and accuracies both on a 0-100 scale; higher is better.
Metric: Morphology Score (0-100; mean of two judge-scored description tasks and two attribute MCQA accuracies). Source: arxiv.org. Saturation forecast: Around November 2028. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 49.86 | #452 |
| 2 | Qwen 2.5 VL 72B Instruct | 47.69 | #845 |
| 3 | GPT-4o Mini | 47.16 | #1627 |
| 4 | Qwen 3 VL 8B Instruct | 45.43 | #982 |
| 5 | Claude Sonnet 4.5 (Thinking) | 45.4 | #244 (Claude Sonnet 4.5) |
| 6 | GLM-4.5V | 44.28 | #739 |
| 7 | QVQ-72B-Preview | 43.09 | #1173 |
| 8 | Llama 3.2 90B | 39.41 | #1740 |
| 9 | Nemotron Nano 12B v2 VL | 36.49 | #2248 |
| 10 | Llama 3.2 11B | 29.97 | #6434 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dermobench-morphology · How It Works · Data refreshed daily, snapshot 2026-11-12.