External Validation of an Artificial Intelligence Algorithm Using Biparametric MRI and Its Simulated Integration with Conventional PI-RADS for Prostate Cancer Detection.
Authors
Affiliations (6)
Affiliations (6)
- Department of Urology, University of Arkansas for Medical Sciences, Little Rock, Arkansas (M.J.B., V.M., J.L.M., E.B., T.L., A.M.A.).
- Department of Radiology, University of Arkansas for Medical Sciences, Little Rock, Arkansas (R.R., J.J., S.B., S.S.).
- Department of Pathology, University of Arkansas for Medical Sciences, Little Rock, Arkansas (N.G.).
- Artificial Intelligence Resource, National Cancer Institute, National Institutes of Health, Bethesda, Maryland (S.A.H.,).
- Molecular Imaging Branch, National Cancer Institute, National Institutes of Health, Bethesda, Maryland (B.T.).
- Department of Urology, University of Arkansas for Medical Sciences, Little Rock, Arkansas (M.J.B., V.M., J.L.M., E.B., T.L., A.M.A.). Electronic address: [email protected].
Abstract
Prostate imaging reporting and data systems (PI-RADS) experiences considerable variability in inter-reader performance. Artificial Intelligence (AI) algorithms were suggested to provide comparable performance to PI-RADS for assessing prostate cancer (PCa) risk, albeit tested in highly selected cohorts. This study aimed to assess an AI algorithm for PCa detection in a clinical practice setting and simulate integration of the AI model with PI-RADS for assessment of indeterminate PI-RADS 3 lesions. This retrospective cohort study externally validated a biparametric MRI-based AI model for PCa detection in a consecutive cohort of patients who underwent prostate MRI and subsequently targeted and systematic prostate biopsy at a urology clinic between January 2022 and March 2024. Radiologist interpretations followed PI-RADS v2.1, and biopsies were conducted per PI-RADS scores. The previously developed AI model provided lesion segmentations and cancer probability maps which were compared to biopsy results. Additionally, we conducted a simulation to adjust biopsy thresholds for index PI-RADS category 3 studies, where AI predictions within these studies upgraded them to PI-RADS category 4. Among 144 patients with a median age of 70 years and PSA density of 0.17ng/mL/cc, AI's sensitivity for detection of PCa (86.6%) and clinically significant PCa (csPCa, 88.4%) was comparable to radiologists (85.7%, p=0.84, and 89.5%, p=0.80, respectively). The simulation combining radiologist and AI evaluations improved clinically significant PCa sensitivity by 5.8% (p=0.025). The combination of AI, PI-RADS and PSA density provided the best diagnostic performance for csPCa (area under the curve [AUC]=0.76). The AI algorithm demonstrated comparable PCa detection rates to PI-RADS. The combination of AI with radiologist interpretation improved sensitivity and could be instrumental in assessment of low-risk and indeterminate PI-RADS lesions. The role of AI in PCa screening remains to be further elucidated.