Using Deep Learning to Inform Differential Diagnoses of Skin Diseases

Abstraction: Deep learning system matches dermatologist accuracy on 26 skin conditions

Key points:

• Google DLS (Deep Learning System) trained on 17,777 teledermatology cases (2010-2018) using Inception-v4 plus up to 45 metadata features (age, sex, symptoms, etc.) to produce ranked differential diagnoses across 26 skin conditions
• Achieved 71% top-1 and 93% top-3 accuracy against a 3-dermatologist voting reference standard (Validation set A, 3,756 cases)
• Top-3 accuracy on Validation set B: DLS 90% vs dermatologists 90% vs PCPs 75% vs NPs 55%
• Performance stable across Fitzpatrick skin types II-IV (top-1: 69-72%; top-3: 91-94%); saliency maps confirmed focus on lesions not skin tone
• Accuracy improves with more images and metadata; absence of metadata degrades performance, partially recoverable by training images-only model
• Published in Nature Medicine (May 2020); ~1.9 billion people affected by skin conditions globally, most seen by non-dermatologists

Connections: Google · Deep Learning · Medical AI · Computer Vision

Source: https://ai.googleblog.com/2019/09/using-deep-learning-to-inform.html?m=1