Multi-institutional Clinical Tool for Predicting High-risk Lesions on 3Tesla Multiparametric Prostate Magnetic...

Area of Science:

• Urology
• Radiology
• Machine Learning

Background:

• Multiparametric magnetic resonance imaging (mpMRI) for prostate cancer detection can lead to overuse of resources and increased costs without careful patient selection.
• There is a need for tools to identify patients most likely to benefit from mpMRI.

Purpose of the Study:

• To develop and validate a clinical tool for predicting the presence of high-risk prostate cancer lesions on mpMRI.
• To improve patient selection for mpMRI, thereby optimizing resource utilization.

Main Methods:

• A support vector machine model was developed using data from 1269 biopsy-naive, prior negative biopsy, and active surveillance patients across four tertiary care centers.
• The model utilized patient age, prostate-specific antigen (PSA), and prostate volume to predict the probability of harboring Prostate Imaging Reporting and Data System (PI-RADS) 4 or 5 lesions.
• Prospective external validation was performed on 214 consecutive patients at two separate institutions.

Main Results:

• Internal validation in 811 patients showed an area under the curve (AUC) of 0.730, with excellent calibration and high net clinical benefit.
• Prospective external validation in two cohorts (n=88 and n=126) demonstrated strong performance with AUCs of 0.740 and 0.744, respectively.
• The final model is available on the Microsoft Azure Machine Learning platform and requires prostate volume as input.

Conclusions:

• The developed BiRCH models can effectively select patients for mpMRI, particularly those who are biopsy-naïve or have had prior negative biopsies.
• This validated calculator aids healthcare professionals and patients in making informed decisions about undergoing prostate MRI.