Mdprostate AI and radiologist performance and interobserver agreement using bpMRI to predict the likelihood of prostate cancer based on PI-RADS

Eduardo Valencia-Segura, Department of Magnetic Resonance Imaging, Hospital Angeles Lomas, Huixquilucan, State of Mexico, Mexico;
Beatriz Elias-Perez, Department of Magnetic Resonance Imaging, Hospital Angeles Lomas, Huixquilucan, State of Mexico, Mexico;
Arturo Hernandez-Medina, Department of Magnetic Resonance, Hospital Angeles Lomas, Huixquilucan, State of Mexico; Biomedical Engineering Area, Faculty of Engineering, Universidad Autonoma de Queretaro, Queretaro, Qro. Mexico;
Francisco J. Omaña-Villagran, Department of Magnetic Resonance Imaging, Hospital Angeles Lomas, Huixquilucan, State of Mexico, Mexico;

Introduction: Artificial intelligence (AI) has improved the diagnostic performance of prostate magnetic resonance imaging (MRI) and reduced interobserver variability in Prostate Imaging Reporting and Data System (PI-RADS) scores. This study compared the diagnostic performance of a radiologist and mdprostate AI, using biparametric (bp) MRI, for predicting the likelihood of clinically significant prostate cancer (csPCa) based on PI-RADS scores and assessed interobserver agreement. Materials and methods: This retrospective cross-sectional study included patients who underwent bpMRI for suspected prostate cancer. bpMRI scans were analyzed using PI-RADS v2.1 by an experienced radiologist and mdprostate. Prostate lesions were classified as low (PI-RADS 1 to 3) or high/very high (PI-RADS 4 to 5) likelihood of csPCa. Sensitivity, specificity, positive and negative predictive values, likelihood ratios, and accuracy were calculated. Cohen’s kappa coefficient was used to assess interobserver agreement, and Bowker’s test of symmetry was used to analyze systematic differences in ordinal PI-RADS scoring. Results: Eighty-two men with a mean age of 64 ± 9.9 years (range 43-85) were included. The mean PSA was 9.9 ± 10.3 ng/mL. Mdprostate showed a sensitivity of 96.3%, specificity of 89.1%, and overall accuracy of 91.5% in predicting high and very high likelihood of csPCa, with the radiologist as the gold standard. Interobserver agreement between the radiologist and mdprostate was almost perfect (k = 0.81; 95% CI, 0.67-0.96). Bowker’s test showed significant differences in PI-RADS categories (p = 0.017), indicating that mdprostate tended to assign higher PI-RADS scores than the radiologist, particularly for PI-RADS 5 lesions. Conclusion: Mdprostate demonstrated high diagnostic performance compared to the radiologist and near-perfect agreement between both, using bpMRI to predict a high or very high likelihood of csPCa based on PI-RADS assessment.

Keywords: Prostate cancer. Artificial intelligence. Biparametric magnetic resonance imaging. Mdprostate. Prostate Imaging. Reporting and Data System.