Esteva 2017: a deep neural network classifies skin cancer at dermatologist level
A single image-recognition network, trained on about 130,000 clinical photographs, told cancerous skin lesions from benign ones as accurately as 21 dermatologists, the first widely cited demonstration that deep learning could match specialists at a cancer diagnosis task.
Overview
Esteva and colleagues at Stanford fine-tuned a convolutional neural network pretrained on everyday images using 129,450 clinical photographs spanning 2,032 skin diseases arranged in a taxonomy. On held-out biopsy-proven images they tested it against 21 board-certified dermatologists on two decisions: keratinocyte carcinomas versus benign seborrheic keratoses, and malignant melanomas versus benign naevi, using both clinical photographs and dermoscopy images. The network's sensitivity and specificity curve matched or exceeded the average dermatologist on each task.
- Training set of 129,450 clinical images covering 2,032 diseases; the network was pretrained on general images and fine-tuned on skin.
- Tested against 21 dermatologists on biopsy-proven images for two binary decisions: keratinocyte carcinoma versus seborrheic keratosis and melanoma versus benign naevus, with and without dermoscopy.
- Performance on both tasks was on a par with the dermatologists across the sensitivity and specificity trade-off.
This paper made AI-assisted skin cancer triage a serious clinical prospect and became the template for later work in radiology and pathology. Prospective trials, regulatory clearance and performance across skin tones followed, and are where its promise is now being tested.
- A retrospective test on curated images, not a prospective clinical study.
- Images came largely from lighter-skinned patients, so performance across skin tones was not established.
- Dermatologists in practice use history and examination, not a photograph alone.
Similar pages
not linked directly; found by shared links- TechnologySkin cancer screening (visual skin examination)
Shares Dermoscopy, total-body photography & AI skin analysis, Basal cell carcinoma, Cutaneous squamous cell carcinoma, Melanoma.
- TechnologyBlack salve and other escharotic pastes
Shares Basal cell carcinoma, Cutaneous squamous cell carcinoma, Melanoma.
- PersonSusana Puig
Shares Basal cell carcinoma, Melanoma.
- CompanyPhilogen S.p.A.
Shares Basal cell carcinoma, Melanoma.
- TreatmentImiquimod
Shares Basal cell carcinoma, Cutaneous squamous cell carcinoma.
- InstitutionUniversity of Arizona Cancer Center
Shares Basal cell carcinoma, Cutaneous squamous cell carcinoma, Melanoma.
- TermBreslow thickness
Shares Dermoscopy, total-body photography & AI skin analysis, Melanoma.
- PersonDirk Schadendorf