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

Abstract

Background: Multiparametric magnetic resonance imaging (mpMRI) for prostate cancer detection without careful patient selection may lead to excessive resource utilization and costs.

Objective: To develop and validate a clinical tool for predicting the presence of high-risk lesions on mpMRI.

Design, setting, and participants: Four tertiary care centers were included in this retrospective and prospective study (BiRCH Study Collaborative). Statistical models were generated using 1269 biopsy-naive, prior negative biopsy, and active surveillance patients who underwent mpMRI. Using age, prostate-specific antigen, and prostate volume, a support vector machine model was developed for predicting the probability of harboring Prostate Imaging Reporting and Data System 4 or 5 lesions. The accuracy of future predictions was then prospectively assessed in 214 consecutive patients.

Outcome measurements and statistical analysis: Receiver operating characteristic, calibration, and decision curves were generated to assess model performance.

Results and limitations: For biopsy-naïve and prior negative biopsy patients (n=811), the area under the curve (AUC) was 0.730 on internal validation. Excellent calibration and high net clinical benefit were observed. On prospective external validation at two separate institutions (n=88 and n=126), the machine learning model discriminated with AUCs of 0.740 and 0.744, respectively. The final model was developed on the Microsoft Azure Machine Learning platform (birch.azurewebsites.net). This model requires a prostate volume measurement as input.

Conclusions: In patients who are naïve to biopsy or those with a prior negative biopsy, BiRCH models can be used to select patients for mpMRI.

Patient summary: In this multicenter study, we developed and prospectively validated a calculator that can be used to predict prostate magnetic resonance imaging (MRI) results using patient age, prostate-specific antigen, and prostate volume as input. This tool can aid health care professionals and patients to make an informed decision regarding whether to get an MRI.

Keywords: Early detection of cancer; Machine learning; Magnetic resonance imaging; Prostatic neoplasm.