Prospective validation of an AI software for detecting clinically significant prostate cancer on biparametric MRI

TL;DR

A new AI software was tested for detecting significant prostate cancer on MRI scans and showed potential to reduce unnecessary biopsies.

Contribution

A fully automated radiomics software was prospectively validated for detecting clinically significant prostate cancer on biparametric MRI.

Findings

The AI software demonstrated feasibility with a 7% failure rate and no serious adverse effects.

The software achieved high diagnostic performance, with potential to reduce unnecessary biopsies.

At an optimized threshold, the software showed higher average precision and fewer false positives per patient compared to the radiologist.

Abstract

To evaluate the feasibility and safety (primary endpoints), and performance (secondary endpoint) of a new artificial intelligence (AI) software for detecting clinically significant prostate cancer (csPCa) on biparametric MRI (bpMRI) compared to an expert radiologist.

In this prospective study at St. Olavs Hospital, Norway (December 2023–October 2024), 89 consecutive biopsy-naïve men underwent bpMRI for suspected PCa. Scans were interpreted by a radiologist using PI-RADS v2.1 and a radiomics-based AI software. Biopsies were obtained from all radiologist- and/or AI-identified lesions. csPCa was defined as ISUP ≥ 2. Feasibility was defined by a < 10% software-failure rate, and safety by the absence of serious adverse device effects (SADEs). Performance was evaluated with ROC, free-response ROC, and precision-recall curves.

Among 89 patients eligible for primary endpoints evaluation, the software demonstrated feasibility (7% failure rate) and safety (no SADEs). Among 76 eligible for secondary endpoint evaluation (median age 68 years [IQR: 63–73]), csPCa was found in 51% (39/76). Patient-level, software achieved an area under the ROC curve [95% CI] of 0.90 [0.83, 0.96] versus 0.86 [0.76, 0.93] (p = 0.25). At a retrospectively optimized threshold matching the radiologist’s patient-level sensitivity at PI-RADS 3 (0.92), software achieved specificity of 0.68 [0.57, 0.78] versus 0.57 [0.46, 0.68] (p = 0.29). Lesion-level, software achieved higher average precision (0.61 [0.52, 0.71] vs. 0.56 [0.46, 0.67]) and lower average false-positive per patient (0.33 [0.22, 0.43] vs. 0.41 [0.30, 0.52]) at the optimized threshold.

The software was feasible and safe, and diagnostic performance showed potential to reduce unnecessary biopsies.

This clinically validated artificial intelligence software enables feasible and safe detection of clinically significant prostate cancer on biparametric MRI, with demonstrated potential to reduce unnecessary biopsies and improve diagnostic accuracy, indicating potential for integration into clinical prostate cancer care.

A fully automated radiomics software for clinically significant prostate cancer detection on biparametric MRI was prospectively clinically validated.The software demonstrated feasibility and safety, with potential to reduce unnecessary biopsies and improve diagnostic accuracy.The investigated radiomics software has the potential for integration into clinical prostate cancer care.

A fully automated radiomics software for clinically significant prostate cancer detection on biparametric MRI was prospectively clinically validated.

The software demonstrated feasibility and safety, with potential to reduce unnecessary biopsies and improve diagnostic accuracy.

The investigated radiomics software has the potential for integration into clinical prostate cancer care.

Figures