Validation of AI-Enhanced MRI Detection of Clinically Significant Prostate Cancer
Overview
This multi-centre UK study validated a CE-certified deep-learning CAD system (Prostate Intelligence™-Pi-v2.4) for detecting clinically significant prostate cancer (csPCa, GG ≥ 2) using MRI data from diverse scanners and hospitals. The AI system's diagnostic accuracy was compared to expert radiologists, demonstrating generalisability across different sites and imaging protocols.
Background
Magnetic resonance imaging (MRI) is critical for early detection and pre-biopsy assessment of prostate cancer, improving patient outcomes by enabling timely intervention. The PI-RADS scoring system standardizes MRI interpretation but variability remains due to differences in radiologist expertise and scanner technology. Deep-learning computer-aided detection (DL-CAD) systems have shown promise in matching expert performance for csPCa detection, but prior studies were mostly single-site and single-scanner, limiting generalisability. Multi-centre validation is essential for clinical translation.
Data Highlights
Parameter
Value
Number of validation patients
252 (42 per site across 6 UK hospitals)
Inclusion criteria
Patients ≥21 years referred for prostate MRI with biopsy or negative MRI report concordance
AI development dataset size
841 cases from 5 sites + 204 cases from PROSTATEx public dataset
Primary endpoint
ROC AUC for detecting GG ≥ 2 prostate cancer
Non-inferiority margin
10%
Key Findings
The AI system (Prostate Intelligence™-Pi-v2.4) was validated on a multi-centre, multi-scanner dataset comprising 252 patients from 6 UK NHS hospitals.
Expert radiologists had >1000 prostate MRI cases experience and provided MDT-supported interpretations as the reference standard.
The AI model was trained on earlier cases from the same sites plus the PROSTATEx dataset, ensuring diverse scanner and protocol representation.
Lesions with GG ≥ 2 were manually annotated and independently verified by expert radiologists, ensuring high-quality ground truth.
The AI system outputs continuous risk scores (1–5) for csPCa without using clinical metadata such as PSA or age.
The primary analysis compared ROC AUCs for csPCa detection between AI and radiologists with a pre-specified 10% non-inferiority margin.
Clinical Implications
This study supports the use of AI-enhanced MRI interpretation to assist radiologists in detecting clinically significant prostate cancer across diverse clinical settings. The demonstrated generalisability across multiple centres and scanner types suggests potential for wider clinical adoption, which may improve diagnostic consistency and reduce unnecessary biopsies. Integration of such AI tools could enhance early detection and patient management.
Conclusion
The validated AI system shows robust performance comparable to expert radiologists in detecting clinically significant prostate cancer across multiple UK centres and scanner types, supporting its clinical utility and generalisability.
References
Sushentsev et al 2022 -- Systematic review of AI algorithms for csPCa detection
PROSTATEx Challenge 2017 -- Public dataset for prostate MRI
Hanley and McNeil 1982 -- Method for comparing ROC AUCs
by Francesco Giganti, Nadia Moreira da Silva, Michael Yeung, Lucy Davies, Amy Frary, Mirjana Ferrer Rodriguez, Nikita Sushentsev, Nicholas Ashley, Adrian Andreou, Alison Bradley, Chris Wilson, Giles Maskell, Giorgio Brembilla, Iztok Caglic, Jakub Suchánek, Jobie Budd, Zobair Arya, Jonathan Aning, John Hayes, Mark De Bono, Nikhil Vasdev, Nimalan Sanmugalingam, Paul Burn, Raj Persad, Ramona Woitek, Richard Hindley, Sidath Liyanage, Sophie Squire, Tristan Barrett, Steffi Barwick, Mark Hinton, Anwar R. Padhani, Antony Rix, Aarti Shah, Evis Sala
Joint clinical consensus outlines evaluation and management considerations for arrhythmias, coronary atherosclerosis, aortic dilatation, myocardial fibrosis, and related findings in older competitive athletes.
