10-Year-Old CNN Outperforms Frontier LLMs in Skin Cancer Diagnosis

A new dermatology AI evaluation led by Dr Magnus Lynch, a consultant dermatologist at Guy’s Hospital and honorary reader at King’s College London, found that a modest convolutional neural network first published in 2015 outperformed two frontier large language models on biopsy-confirmed skin-cancer diagnosis. The assessment, conducted using models available in October 2026, tested GPT-6 Astra and Claude Fable 5.1 against a fine-tuned Inception v3 model using public dermatological images. The benchmark drew on 200 histopathology-confirmed lesions from the MILK10k dataset, released in 2025 through the International Skin Imaging Collaboration. Each lesion included a clinical close-up photograph and a dermoscopic image captured with magnification and illumination. The models were asked to select from six possible diagnoses: melanoma, basal cell carcinoma, squamous cell carcinoma, actinic keratosis, naevus and seborrhoeic keratosis. The results favoured the older CNN. Fine-tuned for roughly twenty minutes on a single H100 GPU using public ISIC Archive images, Inception v3 assigned the correct diagnosis for 125 of 200 lesions, or 62 percent. GPT-6 Astra and Claude Fable 5.1 were correct for 84 and 85 lesions respectively, both reaching 42 percent. When second-choice diagnoses were accepted, the CNN scored 173 correct cases, compared with 134 and 138 for the two large models. The performance gap was notable on several lesion types. The CNN correctly identified 13 of 30 melanomas as its first choice, compared with 8 for GPT-6 Astra and 11 for Claude Fable 5.1. It also performed better on several benign lesions, recognising 37 of 50 moles and 26 of 50 seborrhoeic keratoses. GPT-6 Astra identified 16 moles and 13 seborrhoeic keratoses, while Claude Fable 5.1 identified 33 moles and only 6 seborrhoeic keratoses. The large models also showed distinct failure patterns. GPT-6 Astra overdiagnosed basal cell carcinoma, calling 75 lesions basal cell carcinoma when only 25 were, including 14 melanomas. Claude Fable 5.1 overdiagnosed actinic keratosis, calling 53 lesions actinic keratosis when 20 were. In one illustrative case, a thin invasive melanoma was misidentified by both LLMs as a seborrhoeic keratosis and judged safe for discharge, while the CNN identified it as melanoma. The evaluation suggests that current frontier multimodal models do not necessarily transfer their general visual reasoning to narrow medical image classification tasks. The CNN, although built on an outdated architecture, benefited from task-specific training, dermoscopic image inputs and a tightly controlled evaluation protocol. Images from the same lesion or patient were excluded before training, and the model was not tuned using the test set. The finding has practical implications for clinical AI adoption. It challenges the assumption that larger multimodal systems are automatically superior for medical image interpretation. Instead, architecture fit, training data relevance and evaluation design can make a smaller, specialised network more reliable in a constrained diagnostic setting. The CNN is small enough to run entirely in a mobile phone browser using WebGPU, without uploading patient images. The results were made publicly reproducible with prompts, model responses, scoring code and a downloadable version of the CNN.