AI Measures Prostate Size More Consistently Than Doctors in Cancer Screening
Two AI systems outperformed radiologists at measuring prostate volume on MRI scans, with implications for how clinicians decide who needs biopsy. The finding, from 303 Swedish screening exams, suggests automated tools could standardize a key decision point in prostate cancer detection—potentially reducing unnecessary procedures and improving screening efficiency.
Originaltitel: AI-based prostate volume on MRI in screening of 50-year-old men and its impact on PSA density–based biopsy decisions
<h2>Abstract</h2><h3>Purpose</h3> To compare MRI-based prostate volume measurements by radiologists using the ellipsoid formula with two CE-marked AI models in a population-based screening cohort of 50-year-old men, and to evaluate the impact on PSAD-based biopsy recommendations. <h3>Methods</h3> This retrospective method-agreement study included 303 prostate MRI examinations from Swedish organised prostate cancer testing programmes between 2020 and 2022. Examinations were performed at 14 radiology departments using scanners from multiple vendors and field strengths. Prostate volumes were measured by local radiologists using the PI-RADS–recommended ellipsoid formula and by two CE-marked AI models based on automated whole-gland segmentation. Agreement for prostate volume and PSAD was assessed using Bland–Altman analysis. Biopsy recommendations were evaluated in men with PI-RADS 1–3 lesions using clinically relevant PSAD thresholds. <h3>Results</h3> The mean volume differences between radiologists and AI models were < 2ml. The variability (limits of agreement) between the AI models was lower than that between the radiologist and each AI model. The proportions over the PSAD thresholds were 59–68% for 0.10, 21-28% for 0.15, and 8.6-13% for 0.20ng/ml/cm3 depending on the volumes estimated by the AI models and the radiologists. These differences were significant only for the PSAD threshold 0.15ng/ml/cm3 in combination with PI-RADS ≤3: 29% for the radiologists versus 17% for AI model 2 (p<0.001). <h3>Conclusions</h3> In a screening population of 50-year-old men, CE-marked AI models demonstrated good agreement with radiologists for MRI-based prostate volume estimation, with comparable PSAD-based biopsy recommendations. AI-based volume assessment may reduce variability and radiologist workload in prostate cancer screening.