A decade-old image network beats GPT-6 Astra and Claude Fable 5.1 at diagnosing skin lesions
In a King's College London dermatologist's test on 200 biopsy-proven cases, a fine-tuned Inception v3 named the right diagnosis 62% of the time. Both frontier models reached 42%.
Dr Magnus Lynch, a dermatologist and principal investigator at a King's College London lab that researches AI for skin cancer, compared two leading general-purpose models, GPT-6 Astra and Claude Fable 5.1, with Inception v3. Inception v3 is an image-recognition convolutional neural network that Google released about ten years ago. Lynch fine-tuned it for roughly 20 minutes on 68,011 public dermatology images. He then gave all three models 200 biopsy-proven lesions: 100 that needed examination or treatment and 100 that could be safely discharged. Each model had to pick the most likely diagnosis from six options. The CNN was right in 125 cases (62%). GPT-6 Astra was right in 84 and Claude Fable 5.1 in 85, about 42% each. In one example Lynch highlights, both language models called a thin invasive melanoma on an ear a harmless seborrhoeic keratosis and said the patient could be discharged. The CNN identified it as melanoma from the dermoscopic image. The test is small and was written up as a blog post, not a peer-reviewed study. Still, it suggests that patients and clinicians who turn to general chatbots for skin checks may get less reliable answers than they would from narrow, purpose-trained tools.