Computer Vision
Oct 25, 2026
Older CNN model outperforms advanced AI in skin cancer diagnosis
Oct 25, 2026
AI Summary
A study found that a 10-year-old convolutional neural network (CNN) outperformed two advanced AI models, GPT-6 Astra and Claude Fable 5.1, in diagnosing skin cancer. The CNN correctly identified 62% of lesions compared to 42% for the newer models, highlighting the potential effectiveness of older AI technology in medical applications.
- A dermatologist tested the diagnostic capabilities of GPT-6 Astra and Claude Fable 5.1 against a fine-tuned version of Inception v3, a CNN developed in 2015.
- The study involved 200 biopsy-proven skin lesions, where the CNN achieved a correct diagnosis rate of 62%, while the two AI models scored 42%.
- The CNN was fine-tuned on a dataset of clinical and dermoscopic images, ensuring it had not seen the test lesions before.
- The CNN performed particularly well on benign lesions, identifying 37 out of 50 moles and 26 out of 50 seborrhoeic keratoses correctly, compared to lower rates for the AI models.
- Each AI model had specific biases in their diagnoses, with GPT-6 Astra frequently misidentifying melanomas as basal cell carcinomas and Claude Fable 5.1 misclassifying many lesions as actinic keratosis.
- When asked if patients needed to see a doctor, both AI models recommended medical attention for nearly all cases, including benign lesions.
- The results suggest that older AI models can still be highly effective in specialized medical tasks, such as skin cancer diagnosis.
cnnskin cancerdiagnosishealthcaremedical imaging