The Pulse
Inception v3 Beats Frontier Models in a 200-Lesion Skin Test
Dermatologist Magnus Lynch reports that a fine-tuned Inception v3 CNN correctly classified 62% of 200 biopsy-confirmed skin lesions, compared with 42% for GPT-6 Astra and Claude Fable 5.1. The small test also found a safety trade-off: at it
AI.info Team · Written with AI · fact-checked ·
On October 25, dermatologist Magnus Lynch published a head-to-head test in which a fine-tuned Inception v3 image classifier correctly named 125 of 200 biopsy-confirmed skin lesions. GPT-6 Astra and Claude Fable 5.1 each identified about 42%: 84 and 85 cases, respectively. The CNN, an architecture Google first published in 2015, scored 62% after about 20 minutes of fine-tuning, Lynch writes. But the overall accuracy lead does not translate into a simple safety ranking: at its preset threshold, the CNN would have discharged 18 melanomas among the 30 in the test.
A diagnosis test, not a clinical trial
Lynch drew the cases from MILK10k, a public dataset of lesions photographed both as clinical close-ups and through a dermatoscope. All 200 lesions had histopathology confirmation, and the test set was divided into 100 cases classed as unsafe to discharge and 100 considered safe. The unsafe group contained 30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas and 20 actinic keratoses; the safe group contained 50 moles and 50 seborrhoeic keratoses.
Each language model received both images and the same default prompt, asking for a most likely diagnosis from six options, a confidence level, a second choice and a discharge recommendation. Lynch fine-tuned all layers of Inception v3 on 68,011 clinical and dermoscopic images from the ISIC Archive using one H100 GPU. He says the test lesions and other images from the same lesions or patients were excluded from training and validation, and were not used to select the model or its threshold.
Every lesion in the sample had been biopsied, including the benign ones. Lynch notes that this selection makes the benign cases harder than ordinary examples: a mole removed for biopsy is one that had already concerned a clinician. The results therefore describe this particular set of selected cases, not the mix of lesions seen in routine care.
The accuracy lead hides a safety trade-off
For a first-choice diagnosis, Inception v3 got 13 of 30 melanomas right, compared with eight for GPT-6 Astra and 11 for Claude Fable 5.1. Across all diagnoses, allowing a second choice raised the CNN’s correct count to 173, against 134 and 138 for the language models. The CNN performed better on most categories, but it trailed both models on actinic keratosis and missed one more basal cell carcinoma than GPT-6 Astra.
The discharge figures complicate the headline result. At the threshold Lynch used, the CNN would have kept 77 of the 100 lesions deemed unsafe but discharged 18 melanomas; GPT-6 Astra kept 97 unsafe lesions and discharged three melanomas, while Claude Fable 5.1 kept 96 and discharged four. When Lynch matched the CNN’s operating point to retain 98 or 99 unsafe lesions, it discharged one or two melanomas. The comparison shows why diagnostic accuracy alone cannot establish whether a system is safe to use for clinical triage.
“A ten-year-old network, fine-tuned for about twenty minutes, beat both of them at recognising skin cancer.”
Dr Magnus Lynch, Consultant Dermatologist and Mohs Surgeon at Guy’s Hospital
Lynch flags the limits—and his own role
Lynch calls Inception v3 a baseline, not a state-of-the-art skin cancer system. He says the sample of 200 is too small to rank close results confidently, especially for melanoma, where the test included only 30 cases; each language-model case was run once, and answers can vary between runs. Because MILK10k has been public since 2025, he also says the language models may have seen its images during training, while the CNN had not.
The author says more advanced diagnostic systems use larger image sets and newer architectures, and that regulated products built on them perform better. He also discloses that Claude Code assisted with the analysis and first drafts; Claude Fable 5.1, one of the two models tested, is an Anthropic model. Lynch has posted the prompts, model replies, cases and scoring code, and says the CNN can run in a phone browser without uploading images.
The test makes a narrow claim: on this biopsy-confirmed sample and this six-way classification task, a fine-tuned image classifier outscored two general-purpose models on first-choice diagnosis. Its own discharge results show why that score cannot stand in for evidence of clinical safety.