Dermatologist-level classification of skin cancer with deep neural networks - Esteva et al., Nature, Jan 2017

Original article: https://www.nature.com/articles/nature21056

NOTE: This is one of the earliest, and most cited, articles exploring the use of machine learning in medicine.

One-sentence summary

A group from Stanford trained a neural network to distinguish between certain benign and cancerous skin lesions at the level of performance of board-certified dermatologists.

What did they do?

They trained a convolutional neural network (CNN) to distinguish benign naevi from melanomas (the deadliest skin cancer) and benign seborrheic keratoses from keratinocyte carcinomas (the most common skin cancer) with high accuracy. The 'ground truth' was established by biopsy samples.

The used a larger dataset than any previous studies with a novel CNN architecture (GoogleNet Inception v3) to achieve good performance. They compared the performance with 21 board-certified dermatologists and found the algorithm was broadly able to match or exceed their performance.

What does this mean?

This could increase accessibility to dermatology input, with the potential even for people to check their skin lesions on their phones. It could also act as a decision support tool for clinicians, such as dermatologists or GPs/family doctors, who review many such lesions. The study only focussed on four different lesions - similar approaches could be taken for distinguishing other skin lesions.