A network tells skin cancers apart like dermatologists
On 25 January 2017 Nature published a Stanford convolutional network trained on 129,450 clinical images of 2,032 skin diseases: on two tasks, carcinomas against benign keratoses and melanomas against naevi, it matched 21 board-certified dermatologists.
Why it matters
A network pretrained on ordinary photographs reached specialist level on images taken with a camera rather than a dedicated instrument. The authors conclude that diagnosis could move beyond the clinic to a phone.
The dataset was two orders of magnitude larger than earlier ones; the architecture was Google's Inception v3, pretrained on ImageNet and fine-tuned across all layers. The comparison with dermatologists was on biopsy-proven images. "Dermatologist-level" is the paper's phrase: the network saw images, not patients. A correction followed (Nature 546:686). The forecast of 6.3 billion smartphones the paper cites is someone else's figure, and the record does not repeat it.