Dermatologist: a 10-year-old CNN beats frontier LLMs at skin cancer diagnosis, 62% vs 42%
RazRazcle · x · 2026-10-26
Dermatologist Dr Magnus Lynch tested GPT-6 Astra and Claude Fable 5.1 on 200 biopsy-proven lesions from the public MILK10k dataset, comparing them against a fine-tuned Inception v3 CNN from 2015.
- The CNN scored 62% correct (125/200); the frontier LLMs only 42% (84 and 85/200) — after 20 minutes of fine-tuning
- On melanoma: CNN 13/30, Claude 11/30, GPT 8/30. Both LLMs misread an invasive ear melanoma (<0.8mm) as a seborrhoeic keratosis and recommended discharge
- LLMs were strong on basal cell carcinoma: GPT got 25/25
- Takeaway: real-world long-tail domains where precision matters still need domain-specific training data LLMs don't yet have
The poster notes such training data will inevitably reach LLMs — its absence today is the surprise.
More from Models
- Anthropic engineer: Claude Code cache hit rates are now high, savings live — teortaxesTex · 2026-10-26
- Fine-tuned 2015 CNN beats GPT-6 and Claude on skin lesion diagnosis, 62% vs 42% — lmoroney · 2026-10-26
- "Never seen Claude this happy": community speculates Opus 4.5 is imminent — adonis_singh · 2026-10-26
- 'Absolute humiliation': China open-weights model outscoring OpenAI flagship draws mockery — alejandroll10 · 2026-10-26
- Claude's reverse-engineering refusals spark debate; OpenAI says Codex blocks are accidental — sebkrier · 2026-10-26
- Blogger says Claude Opus 5.5 is the best model he's used, well ahead of GPT-6 Astra — AlchainHust · 2026-10-26