Machine learning in medical imaging

Assembled from 21 blocks · 2 evidence · 32 related

1. Story
2. The clinical domain with the longest record of confident prediction and the best randomised evidence, and the two do not agree. In 2016 Geoffrey Hinton publicly advised that training of radiologists should stop, on the view that deep learning would supersede human image reading within about five years. What the evidence since actually shows is narrower and better: in Sweden's MASAI trial, a randomised, population-based screening study of over 100,000 women, AI-supported reading raised the screen-detected cancer detection rate by 29% and cut screen-reading workload by 44.2% against standard double reading, with seven more false positives across the arms; the full results published in The Lancet reported a non-inferior interval-cancer rate, higher sensitivity, unchanged specificity, and fewer interval cancers with unfavourable characteristics. The workforce moved the other way from the prediction: the American College of Radiology's February 2025 workforce study projects growth, and United States employment projections for the occupation exceed the all-occupation average. A tool that reads alongside a radiologist and removes half the reading load is a substantial change; it is not the change that was announced.
3. Knowledge
4. Deep learning
5. Machine learning in medical imaging
6. Connections
7. Deep learning
8. European Union Artificial Intelligence Act
9. Geoffrey Hinton
10. A randomised screening trial reports what AI-supported mammography actually changes
11. The AI accelerator
12. ImageNet
13. The transformer architecture
14. AlphaFold
15. Machine-learned weather forecasting
16. Machine learning in medical imaging
17. A deep convolutional network wins the ImageNet challenge
18. AlphaGo defeats Lee Sedol in Seoul
19. Evidence
20. MASAI is a randomised, controlled, non-inferiority, single-blinded, population-based screening trial in Sweden enrolling over 100,000 women. Reported results include a 29% higher screen-detected cancer detection rate with AI support, a 44.2% reduction in screen-reading workload (48,444 fewer readings), false-positive counts of 772 (1.5%) in the intervention arm against 765 (1.4%) in the control arm, and in the full results a non-inferior interval-cancer rate reported as about 12% lower, higher sensitivity, unchanged specificity and fewer interval cancers with unfavourable characteristics. V55 verification basis: none of the three papers was fetched — thelancet.com and sciencedirect.com are blocked to this session. The figures and the trial design were carried in search results attributing them to the named Lancet titles and to institutional press material about them; the publication dates of the final results are reported variously as late 2025 and early 2026 and were NOT resolved, and no DOI is asserted.
21. An ACR workforce study published in February 2025 using CMS data for 2014–2023 found 37,482 radiologists enrolled to provide care to Medicare patients in 2023 and projected growth of 25.7% to 40.3% by 2055 depending on residency expansion; United States employment in the occupation is projected to grow 5% from 2024 to 2034, above the 3% all-occupation average. In 2016 Geoffrey Hinton publicly advised that training of radiologists should stop, on the expectation that deep learning would supersede human image reading within about five years; he has since acknowledged the prediction was wrong on timing and stated too broadly. V55 verification basis: neither the ACR study nor the BLS projection page was fetched. The 37,482 count, the projection range, the 5% growth figure and the account of Hinton's remark and his later acknowledgement were carried in search results attributing them to the ACR study, to BLS projections and to contemporary reporting; the wording and occasion of the 2016 remark were NOT confirmed from a primary record.