The 2017 Nature paper “Dermatologist-level classification of skin cancer with deep neural networks,” by Andre Esteva, Sebastian Thrun, and colleagues at Stanford, trained its convolutional neural network on 129,450 clinical images spanning 2,032 different skin diseases. The authors described this dataset as two orders of magnitude larger than those used in previous skin-classification work.
Rather than learn from scratch, the network began as a GoogLeNet Inception v3 model pretrained on roughly 1.28 million general ImageNet photographs, then was fine-tuned on the skin images - an example of transfer learning making a data-hungry medical task feasible. The scale of the labeled dataset was a central reason the model could reach performance on par with board-certified dermatologists on the tested diagnostic tasks.