Validation of a Machine Learning Model Based on Multiparametric MR for the Prediction of Clinically Significant Prostate Cancer

This is a transversal, non-interventional observation study that involves the analysis of data collected both retrospectively and prospectively.

Patients will be enrolled by the radiologist medical staff designated by the Principal Investigator. The Principal Investigator and his delegates will be responsible for the acquisition and pseudo-anonymization of images of patients enrolled (both retrospectively and prospectively) from the RIS-PACS of the IRCCS Bologna University Hospital, Policlinico di Sant'Orsola. When available, for the participants undergoing prostate surgery, the O.U. of Pathological Anatomy will provide digital scans of the prostatic macrosections.

For the validation of the machine learning model of clinically significant CaP, a tool will be made available to radiologists. This tool is developed by the Department of Informatics - Science and Engineering, DISI, University of Bologna and Alma Mater Research Institute on Global Challenges and Climate Change, Alma Climate, University of Bologna, and integrated in an open-source Dicom viewer, Aliza MS Dicom Viewer. The doctors will then have the possibility to carry out independently the segmentation of the ADC sequences (previously pseudo-anonymized) and start the radiomic process to obtain a predictive value of the probability that the segmented lesion is clinically significant. This probability value will be calculated by the software on the basis of the machine learning model developed in PROSTATE_01. The tool also allows data collection (organized and pseudo-anonymous). The validation of the clinically significant predictive model of CaP will be carried out on all patients enrolled both retrospectively and prospectively, excluding those used in the model training phase in PROSTATE_01. After having evaluated the performance of the validated model and the effects of selection bias, once the recruitment is completed both retrospectively and prospectively, a refinement of the previously developed model will be carried out. This will be a "re-training" of the model itself, performed on training and test datasets randomly selected from the entire enrolled population from which it will be obtained randomly, by difference, also the subset for a new validation.