A Pilot Study of AI-Assisted Reading of Prostate MRI in Organized Prostate Cancer Testing

Interest in population-based prostate cancer screening is growing, particularly since the European Council's 2022 recommendation to evaluate organized screening approaches. Sweden, where I practice as a radiologist, is at the forefront of large-scale screening studies—such as the Göteborg-2 study—and of organized testing in clinical routine through the OPT project (Organized Prostate Cancer Testing).

One of the main challenges of implementing a screening program is avoiding the detection of indolent cancers, a major barrier for PSA-based screening programs. Ongoing screening studies incorporate prostate MRI to select men for biopsy, and they also explore additional filters to reduce the number of MRI exams needed. Our current study investigates how available AI models perform in lesion detection and characterization on MRI, as well as the implications of implementing such models. Given the rapid development of AI models for this purpose, our results should be interpreted with this progression in mind.

In this pilot study, we evaluated the only commercially available AI model at the time that offered PIRADS grading. Today, there are several models available, and our research team recently concluded an evaluation of over 400 MRI scans from multiple centers within the OPT framework. The upcoming publication will examine differences between local primary review and the potential added value of expert review, as well as compare the performance of multiple AI models on a larger dataset. In the past year, retrospective studies have shown promising performance for AI models (e.g., the PI-CAI publications); however, there is a lack of relevant studies focused on screening cohorts. We believe that our current and forthcoming studies in this area can address an important gap.

From a patient perspective—or more precisely, from the perspective of men in a screening program who are not yet patients—it is crucial to thoroughly evaluate and balance the advantages and disadvantages of implementing AI models within the diagnostic process. Our findings from the published pilot study indicate that introducing AI could lead to a significant increase in healthy men selected for biopsy, which may result in over-detection and over-diagnosis of indolent cancers. This, in turn, could place many men under prolonged active surveillance, potentially impacting quality of life, or lead to treatments with well-known complications and side effects, in addition to economic costs. We therefore encourage continued research initiatives to explore this area further.

Written by: Erik Thimansson, PhD, Consultant Radiologist, Institution for Translational Medicine, Diagnostic Radiology, Lund University, Skånes Universitetssjukhus Malmö, Malmö, Sweden

Read the Abstract

Login