AI decision support for increasing prostate biopsy efficiency: a retrospective multicentre, multiscanner study.
Authors
Affiliations (19)
Affiliations (19)
- Department of Radiology, Addenbrooke's Hospital and University of Cambridge, Cambridge, UK. [email protected].
- Lucida Medical Ltd, Cambridge, UK.
- Somerset NHS Foundation Trust, Taunton, UK.
- University of Winchester, Winchester, UK.
- Hampshire Hospitals NHS Foundation Trust, Winchester, UK.
- Department of Urology, Hertfordshire and Bedfordshire Urological Cancer Centre, Lister Hospital, Stevenage, UK.
- School of Life and Medical Sciences, University of Hertfordshire, Hatfield, UK.
- East and North Hertfordshire Teaching NHS Trust, Stevenage, UK.
- Royal Cornwall Hospitals NHS Trust, Truro, UK.
- Royal United Hospitals Bath NHS Foundation Trust, Bath, UK.
- Mid and South Essex NHS Foundation Trust, Southend, UK.
- North Bristol NHS Trust, Bristol, UK.
- Division of Surgery and Interventional Science, UCL, London, UK.
- Centre for Urology Imaging, Prostate, AI and Surgical Studies (COMPASS) Research Group, Division of Surgery and Interventional Science, UCL, London, UK.
- Department of Urology, University College London Hospitals NHS Foundation Trust, London, UK.
- Department of Radiology, Addenbrooke's Hospital and University of Cambridge, Cambridge, UK.
- Paul Strickland Scanner Centre, Mount Vernon Hospital, Northwood, UK.
- Dipartimento Diagnostica per Immagini e Radioterapia Oncologica, Policlinico Universitario A. Gemelli IRCCS, Rome, Italy.
- Dipartimento di Scienze Radiologiche ed Ematologiche, Università Cattolica del Sacro Cuore, Rome, Italy.
Abstract
To develop and retrospectively validate an artificial intelligence-based decision support system (AI-DSS) for optimising prostate biopsy decisions and improving benefit-to-harm ratios. This retrospective, multicentre, multiscanner study used data from 1022 patients. An AI-DSS integrating PI-RADS scores, automated prostate-specific antigen density (PSAd), and deep-learning imaging risk scores was developed on 770 cases and validated on an independent cohort of 252 men from six UK centres. The AI-DSS performance was benchmarked against the real-world clinical decisions (reference standard) using grade selectivity, biopsy efficiency, and selective biopsy avoidance as outcome measures. Biopsy-proven detection of grade group (GG) ≥ 2 disease was the reference standard. In the validation cohort of 252 patients (mean age, 67.3 years), 137 underwent biopsy and 79 (31%) harboured ≥ GG2 disease. Compared to the reference standard, the AI-DSS at the 31% cancer detection rate (CDR) would have avoided 28 biopsies while missing one ≥ GG2 cancer. This corresponded to a 70% increase in grade selectivity (from 4.6 to 7.8), 79% increase in biopsy efficiency (from 1.4 to 2.5), and a 143% increase in selective biopsy avoidance (from 2.8 to 6.8). At the reduced CDR of 30%, grade selectivity, biopsy efficiency, and selective biopsy avoidance increased by 172%, 236%, and 475%, with four ≥ GG2 cancers missed. An AI-DSS that integrates clinical and advanced imaging data improves the benefit-to-harm ratio of prostate biopsy decisions in a retrospective setting. Future prospective validation as part of real-world clinical workflow is required to enable clinical implementation. Question Current prostate cancer diagnostic pathways result in fewer unnecessary biopsies. Can an AI decision support system (AI-DSS) further improve biopsy efficiency for detecting significant cancer? Findings An AI-DSS avoided 28 biopsies in a 252-patient cohort, increasing grade selectivity, biopsy efficiency, and selective biopsy avoidance by 70%, 79%, and 143%, respectively. Clinical relevance Integrating an AI-DSS into clinical workflows may further reduce unnecessary prostate biopsies and overdiagnosis of indolent disease, thus potentially improving the efficiency of the prostate cancer diagnostic pathway.