AI detects prostate cancer more frequently than radiologists. Additionally, AI generates false alarms half as often. This is demonstrated by an international study coordinated by the Radboud University Medical Center and published in The Lancet Oncology. This is the first large-scale study in which an international team transparently evaluates and compares AI with radiologist evaluations and clinical outcomes.
Radiologists face an increasing workload as men with a higher risk of prostate cancer now routinely receive an MRI of the prostate. Diagnosing prostate cancer using MRI requires extensive experience and there is a shortage of experienced radiologists. AI can help with these challenges.
AI expert Henkjan Huisman and radiologist Maarten de Rooij, project leaders of the PI-CAI study, organized a major competition between AI teams and radiologists with an international team. Together with other centers in the Netherlands and Norway, they performed more than 10,000 MRI scans. They transparently determined for each patient whether prostate cancer was present. They allowed several groups around the world to develop artificial intelligence to analyze these images. The top five submissions were combined into a super algorithm to analyze MRI images for prostate cancer. Finally, the IA’s assessments were compared with those of a group of radiologists on four hundred prostate MRIs.
Accurate diagnosis
The PI-CAI community brought together more than two hundred AI teams and 62 radiologists from twenty countries. They compared the AI and radiologists’ findings not only to each other but also to a gold standard, while monitoring the results of the men from whom the scans originated. On average, the men were followed for five years.
This first international study on AI in prostate diagnosis shows that AI detects almost seven percent more major prostate cancers than radiologists. Additionally, AI identifies suspicious areas that are later found to be non-cancerous, fifty percent less often. This means that the number of biopsies could be reduced by half with the use of AI. If these results are replicated in follow-up studies, they could be of great help to radiologists and patients in the future. It could reduce radiologists’ workload, provide more accurate diagnoses and minimize unnecessary prostate biopsies. The developed AI still needs to be validated and is currently not yet available for patients in clinical settings.
Quality system
Huisman observes that society has little trust in AI. “This is because sometimes manufacturers create AI that is not good enough,” he explains. He is working on two things. The first is a public and transparent test to fairly evaluate AI. The second is a quality management system, similar to that which exists in the aviation industry. ‘If planes almost collide, a safety committee will study how to improve the system so that this does not happen again in the future. I want the same for AI. I want to research and develop a system that learns from every mistake so that the AI is monitored and can continue to improve. That way, we can build trust in AI for healthcare. Optimal, governed AI can help make healthcare better and more efficient.”