Material Insights · Understanding the Material World

AI diagnosis from medical images

Neural networks trained on scans flag disease in images such as retinas and mammograms.

Health and medicine · Diagnosis · No. 198

  • Networked and AI age
  • Readiness 8 of 9
  • System
Health and medicine · Technology

Since 2018 a computer has been allowed to screen eyes for diabetic damage with no doctor looking at the picture.

Chart: Illustration: AI diagnosis from medical images

Where it comes from

Chart: Map of where it comes from

Computer aids for reading scans existed from the 1990s. Deep learning changed them: in 2012 a deep network won a contest on breast cancer tissue images, and in 2016 one graded retinal photographs.

Three things to know
  1. A 2016 network for diabetic eye disease learned from 128,000 retinal images graded by ophthalmologists.
  2. In an early analysis of a Swedish trial of 80,033 women, AI-supported screening found 6.1 cancers per 1,000 women against 5.1 with two human readers.
  3. Most AI devices authorised in the United States read radiology images.
The journey
  1. 1 ImagesHospitals collect scans, photographs and slides
  2. 2 LabelsSpecialists grade each image, often several per image
  3. 3 TrainingA neural network learns the patterns that match each grade
  4. 4 TestingThe software is checked on new images and in trials
  5. 5 ApprovalRegulators authorise it as a medical device
  6. 6 ClinicIt flags, ranks or reads images alongside or instead of a doctor
Go deeper pick a tab to open it in full