ABSTRACT
PURPOSE: To assess the performance of a computer-aided diagnosis (CADx) system trained at characterizing International Society of Urological Pathology (ISUP) grade≥2 peripheral zone (PZ) prostate cancers on multiparametric magnetic resonance imaging (mpMRI) examinations from a different institution and acquired on different scanners than those used for the training database. PATIENTS AND METHODS: Preoperative mpMRIs of 74 men (median age, 65.7 years) treated by prostatectomy between 2014 and 2017 were retrospectively selected. One radiologist outlined suspicious lesions and scored them using Prostate Imaging-Reporting and Data System version 2 (PI-RADSv2); their CADx score was calculated using a classifier trained on an independent database of 106 patients treated by prostatectomy in another institution. The lesions' nature was assessed by comparison with prostatectomy whole-mounts. Diagnostic accuracy was estimated with areas under receiver operating characteristic curves (AUCs). Sensitivity and specificity were calculated using a CADx threshold (≥0.21) that yielded 95% sensitivity in the training database, and a PI-RADSv2≥3 threshold. RESULTS: A total of 127 lesions (PZ, n=104; transition zone [TZ], n=23) were described. In PZ, CADx and PI-RADSv2 scores had similar AUCs for characterizing ISUP grade≥2 cancers (0.78 [95% confidence interval (CI): 0.69-0.87] vs. 0.74 [95%CI: 0.62-0.82], respectively) (P=0.59). Sensitivity and specificity were respectively 89% (95%CI: 82-97%) and 42% (95%CI: 26-58%) for the CADx score, and 97% (95%CI: 93-100%) and 37% (95%CI: 22-52%) for the PI-RADSv2 score. In TZ, both scores showed poor specificity. CONCLUSION: In this external cohort, the CADx and PI-RADSv2 scores showed similar performances in characterizing ISUP grade≥2 cancers.
