Source: Clinical study and open-source evaluation benchmark published by Dr. Magnus Lynch on October 25, 2026.
Dr. Magnus Lynch, Consultant Dermatologist and Mohs Surgeon at Guy's Hospital and Honorary Reader at King's College London, constructed an empirical evaluation utilizing 200 histopathologically confirmed cutaneous lesions from the MILK10k archive. The dataset pairs clinical macroscopic photography with dermoscopic microscopy under illuminated magnification.
The benchmark challenged the systems with 100 lesions requiring clinical intervention (30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas, and 20 actinic keratoses) alongside 100 safe lesions (50 benign naevi and 50 seborrhoeic keratoses). Inception v3 was fine-tuned on 68,011 archival ISIC images across a single NVIDIA H100 GPU in approximately twenty minutes, with all test cases and patient identities strictly sequestered.
When permitting a second diagnostic differential, Inception v3 achieved 173 correct classifications out of 200, compared to 138 for Claude Fable 5.1 and 134 for GPT-6 Astra. While the frontier models retained 96 to 97 unsafe lesions, they suffered from defensive over-calling that misclassified common benign growths as malignant, severely degrading triage specificity.
Notably, the fine-tuned CNN operates with a parameter footprint small enough to execute entirely on-device via WebGPU in a mobile browser without cloud offloading. Both frontier LLMs demonstrated significant vulnerability to prompt sensitivity, defaulting to conservative recommendations to consult physicians when presented with clinical macroscopic images alone.
What remains unexplored is how modern vision-language models would perform if fine-tuned directly on calibrated dermoscopic datasets rather than evaluated zero-shot with general-purpose multimodal weights.