Use of Artificial Intelligence in prostate MRI: A rapid scoping review highlighting limited evidence in screening context.
Authors
Affiliations (9)
Affiliations (9)
- Early Detection, Prevention, and Infections Branch, International Agency for Research on Cancer (IARC/WHO), Lyon, France. Electronic address: [email protected].
- Department of Radiology, Althaia Foundation, Manresa, Spain.
- Department of Urinary and Vascular Imaging, Hôpital Edouard Herriot, Hospices Civils de Lyon, Lyon, France; University Lyon 1, Lyon, France.
- Department of Radiology, University College London Hospital NHS Foundation Trust, London, UK; Division of Surgery & Interventional Science, University College London, London, UK.
- Department of Radiology, Hospital Universitario de Vigo, Spain.
- Department of Urology, Erasmus Cancer Institute, Erasmus University Medical Center Rotterdam, the Netherlands.
- European Association of Urology, Policy Office, Arnhem, the Netherlands.
- European Association of Urology, Policy Office, Arnhem, the Netherlands; Department of Urology, KU Leuven, Belgium.
- Early Detection, Prevention, and Infections Branch, International Agency for Research on Cancer (IARC/WHO), Lyon, France.
Abstract
Artificial Intelligence (AI) is seen as a potential solution to alleviate workforce demands arising from growing use of magnetic resonance imaging (MRI) in prostate cancer (PCa) screening. We aimed to synthesize the evidence on use of AI in prostate MRI readings in asymptomatic men in PCa screening settings. We conducted a rapid scoping review following PRISMA-ScR guidelines and Cochrane rapid review methods performing the systematic search of major databases supplemented by grey literature search with no restrictions in study design and time-duration. We considered various aspects of utilization of AI in MRI interpretations and biopsy indications in the screening setting. Anticipating limited evidence on AI implementation in screening settings, we extended the review from 'what is known' to discussion on 'key considerations for expected expansion'. We identified 284 records with 47 studies assessed for eligibility and two studied met the inclusion criteria. Both evaluated commercially available ProstateAI software tool to interpret prostate MRI. Agreement between deep learning-based algorithm of AI and expert radiologist ranged from poor to moderate (kappa 0.17-0.42). AI demonstrated high tendency of over-detection and low specificity, leading to discordance with expert radiologists. Current evidence on use of AI in prostate MRI interpretation is limited, but this review highlights several important directions for future research and implementation. Generating robust evidence base in the coming years will be crucial to ensure that AI integration enhances the effectiveness and acceptability of future prostate cancer screening programs. In this study, we examined whether Artificial Intelligence (AI) tools can accurately read prostate MRI scans of apparently healthy men for early detection of prostate cancer. We found that deploying current AI tools that are trained and tested in hospital referred patients may not be optimal to read MRI performed in asymptomatic men. We conclude that AI needs much more training and testing in real screening populations before it can be safely used in prostate cancer screening programs.