+ HEALTH BACKGROUND/OBJECTIVES: Cardiovascular diseases (CVDs) remain the leading cause of morbidity and mortality wor…+ LABOR We introduce GDPval, a benchmark evaluating AI model capabilities on realworld economically valuable tasks. G…+ LABOR Labor market forecasting relies heavily on economic time-series data, often overlooking the “health–wealth” g…− ENTERTAINMENT Generative artificial intelligence (AI) has transformed music production from a specialist studio practice in…+ CLIMATE Water holds great significance as a vital resource in our everyday lives, highlighting the important to conti…+ ENTERTAINMENT With YouTube Music becoming one of Google AI Mode’s new Connected Apps, I used these 7 prompts to create new…− ENTERTAINMENT Some musicians and artists in Peterborough are pushing back on the use of artificial intelligence to create c…+ LABOR Earlier this year Paul* discovered via a flurry of media stories that his business group had been targeted fo…
TruaceTracing the truth around AITuesday, July 21, 2026
Subscribe
Health·P Space·Evidence-backed problem·Published 2026-07-20

Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study

BACKGROUND: Artificial intelligence (AI) systems can potentially aid the diagnostic pathway of prostate cancer by alleviating the increasing workload, preventing overdiagnosis, and reducing the dependence on experienced radiologists. We aimed to investigate the performance of AI systems at detecting clinically significant prostate cancer on MRI in comparison with radiologists using the Prostate Imaging-Reporting and Data System version 2.1 (PI-RADS 2.1) and the standard of care in multidisciplinary routine pract…

TRV-2026-0402Peer-reviewedPermanent record — cite & verify
Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study

Costs and effectiveness of prostate cancer screening in elderly men by United States. Congress. Office of Technology Assessment. Public domain

The quick read

Researchers trained an AI system on 9207 prostate MRI examinations from the Netherlands and tested it on 1000 examinations from the Netherlands and Norway, with a 400-case subset read by 62 radiologists from 20 countries. By June 2024 publication, the AI achieved AUROC 0.91 versus 0.86 for radiologists using PI-RADS 2.1, and at matched operating points detected 6.8% more clinically significant cancers at same specificity or 50.4% fewer false positives at same sensitivity.

The result matters because prostate MRI workload is rising and depends on experienced readers; a supportive AI that reduces overdiagnosis of grade group 1 cancers while maintaining sensitivity could aid primary diagnostic settings. Uncertainty remains because the comparison to multidisciplinary routine practice showed slightly lower specificity and the authors state prospective validation is needed to test clinical applicability.

Main points
  • International paired non-inferiority study trained on 9207 MRI examinations from 11 sites in the Netherlands and tested on 1000 examinations from 12 sites in Netherlands and Norway.
  • Multireader study included 62 radiologists from 45 centres in 20 countries with median 7 years prostate MRI experience reading 400 paired examinations.
  • Reference standard used histopathology and at least 3 years median 5 years follow-up; 2440 of 10,207 examinations had Gleason grade group 2 or greater cancer.
  • Against multidisciplinary routine practice in 1000 testing cases, AI showed 68.9% specificity versus 69.0% at 96.1% sensitivity.
Problem

When compared to historical multidisciplinary routine practice readings in 1000 testing cases, the AI system did not demonstrate confirmed non-inferiority and showed slightly lower specificity at matched sensitivity.

The rundown

The study used 10,207 MRI examinations from 9129 patients collected from Jan 1, 2012 through Dec 31, 2021, with 9207 cases for training and tuning and 1000 cases for testing, plus a 400-case multireader subset evaluated by 62 radiologists from 45 centres in 20 countries.

Primary endpoints were sensitivity, specificity and AUROC compared to PI-RADS 2.1 readers and to historical radiology readings made during multidisciplinary routine practice with patient history and peer consultation, using histopathology and at least 3 years follow-up as reference.

Sources

Reader signal

How should this claim be treated?

The debate