A convolutional network fine-tuned from a 2015 design diagnosed skin cancer more accurately than frontier language models: 62% vs 42% on 200 biopsy-confirmed lesions
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
For narrow medical tasks, a small specialized model can still outperform general-purpose ones.
A dermatologist compared GPT-6 Astra and Claude Fable 5.1 with an Inception v3 network, fine-tuned on 68,011 images from the ISIC archive on a single H100 GPU, on 200 lesions from the MILK10k set (clinical photos and dermoscopy; 100 needed treatment, 100 did not). The network gave the correct diagnosis in 62% of cases (125 of 200) and the language models in 42% (84 and 85); for melanoma they found 13 of 30 cases against 8 and 11. The author names limitations: small samples (30 melanomas), all lesions were biopsied (that is, harder cases), possible leakage of the data into the language models' training, and the network is only a baseline, not the best method. This is an independent analysis on a personal website, not a peer-reviewed study.
Sources 1