Artificial Intelligence Software (Prostate Intelligence) for Detecting Clinically Significant Prostate Cancer on MRI

Background and rationale Prostate cancer is the most common cancer in men in Europe. International guidelines recommend multiparametric MRI before biopsy, and findings are graded with PI-RADS v2.1. PI-RADS assessment varies with institution, reader experience and field strength, and inter-reader variability is substantial among less experienced radiologists. Prostate Intelligence (Pi, Lucida Medical Ltd) is an AI software that analyses prostate MRI and provides prostate volume, lesion detection/segmentation, an AI-Likert score and a preliminary report as a concurrent read to support the radiologist. This study evaluates the diagnostic performance of Pi in a prospective clinical setting. Design Prospective, multicentre (national), interventional diagnostic accuracy study with a medical device. The software runs on servers of the participating institutions; DICOM data are de-identified and imaging data do not leave the institutional perimeter. Reading workflow For each examination the radiologist first reports the MRI blinded to Pi; the report is locked and time-stamped in RIS/PACS. The Pi output is then made available and the radiologist produces a merged Pi-assisted report, recording PI-RADS v2.1 categories, lesion zone and whether each Pi finding is accepted or rejected, and adding any lesions considered missed by Pi. Reference standard Histopathology from MRI-targeted biopsy (maximum two lesions, including discordant radiologist/Pi lesions) and, where available, radical prostatectomy. Participants who are not biopsied are followed for 12 months after MRI (PSA at 6 and 12 months plus clinical assessment) and classified as composite true negatives when there is no evidence of clinically significant cancer. Clinically significant prostate cancer (csPCa) is defined as Gleason grade 3+4 or higher. Endpoints Co-primary: patient-level sensitivity and specificity of stand-alone Pi for csPCa (Gleason \>=3+4). Secondary (all exploratory): diagnostic performance of the Pi-assisted read at patient level for Gleason \>=3+4 and \>=4+3; comparison of Pi-assisted vs radiologist-alone and Pi vs radiologist-alone; proportion of csPCa detected and additional csPCa detection with Pi; lesion-level performance of the Pi-assisted read; exploratory cost-consequence analysis; subgroup performance in patients with radical prostatectomy histology. Sample size 226 participants (150 at Gemelli, 76 at San Raffaele), based on 7.5% precision at a 5% significance level for an expected Pi sensitivity of 90% and specificity of 65%, with an assumed csPCa prevalence of 31% (R v4.4.2). Assumptions are conservative relative to published validation of the device. The study is powered for stand-alone accuracy only; comparative endpoints are exploratory. Statistical analysis Sensitivity, specificity, PPV, NPV and ROC AUC with 95% CIs. AUCs of radiologist-alone, Pi-alone and Pi-assisted reads compared with the DeLong method, with bootstrapped two-sided 95% CIs. Lesion-level metrics (PI-RADS \>=3 as positive) use generalised estimating equations to account for clustering. The primary analysis includes histologically verified participants and composite true negatives; pre-specified sensitivity analyses are restricted to histologically verified participants and a bounding analysis assigning unverified participants first to non-csPCa and then to csPCa. Missing data are tabulated and the primary analysis uses complete cases. Exploratory comparative p-values are reported unadjusted and Benjamini-Hochberg adjusted (5%). Centre, scanner model and field strength are recorded and explored as subgroups; no centre contributes more than 70% of participants. Reporting follows STARD. Duration 27 months: 12 months enrolment and data collection, up to 12 months follow-up per participant, 3 months analysis. Safety All adverse events are recorded; device incidents and deficiencies are reported to the National Competent Authority and the manufacturer under EU Regulation 2017/745 and Italian Legislative Decree 137/2022.