AI Assessment for Prostate Cancer Detection in Biparametric MRI Screening

Overview

This study evaluated a neural network trained on biparametric MRI data from a prostate cancer screening cohort to detect clinically significant prostate cancer (csPC). The AI system was compared against expert radiologist assessments using a rigorous reference standard derived from biopsy and follow-up data. Results highlight the potential of AI to aid in screening workflows by identifying csPC with high accuracy.

Background

Prostate MRI is increasingly recommended before biopsy to improve diagnostic accuracy and reduce unnecessary procedures. However, radiological interpretation is resource-intensive and subject to variability. Artificial intelligence (AI) offers a promising solution to streamline MRI reading, reduce costs, and potentially improve consistency in prostate cancer screening programs. This study addresses the gap in evidence regarding AI performance specifically in screening populations with lower cancer prevalence.

Data Highlights

The study population consisted of men aged 50–60 years from the Göteborg Prostate Cancer Screening 2 Trial who underwent biparametric MRI between 2015 and 2019. MRI data included axial T2-weighted and diffusion-weighted imaging with ADC maps. Reference standards were established using biopsy results and follow-up data, with clinically significant prostate cancer defined as ISUP grade ≥ 2. Three experienced radiologists performed consensus reads for primary MRI assessment using PI-RADS v2 scoring.

Key Findings

• The AI model was trained using the nnU-Net deep learning framework on biparametric MRI sequences (T2W, high b-value DWI, and ADC maps).
• Reference standard determination incorporated multiparametric MRI, biopsy results, and longitudinal follow-up to confirm clinically significant prostate cancer presence.
• AI performance was benchmarked against expert radiologists’ biparametric PI-RADS assessments, focusing on detection of csPC.
• The study population was derived from a randomized screening trial, ensuring relevance to screening settings with lower cancer prevalence compared to clinical cohorts.
• Use of biparametric MRI (excluding dynamic contrast-enhanced sequences) aligns with efforts to simplify and expedite prostate MRI protocols in screening.

Clinical Implications

AI-assisted interpretation of biparametric prostate MRI may reduce radiologist workload and improve screening efficiency by reliably identifying patients with clinically significant prostate cancer. Incorporating AI could help address shortages of expert readers and minimize inter-reader variability. However, further validation in prospective screening cohorts is necessary before widespread clinical implementation.

Conclusion

This study demonstrates the feasibility of training AI models on biparametric MRI data from a prostate cancer screening cohort to detect clinically significant disease. AI has the potential to augment prostate cancer screening workflows, but continued evaluation is essential to confirm clinical benefits and cost-effectiveness.

References

1. Göteborg Prostate Cancer Screening 2 Trial (G2-trial) -- ISRCTN94604465
2. nnU-Net Deep Learning Framework -- Isensee et al. 2021
3. PI-RADS v2 Guidelines -- Weinreb et al. 2016
4. ISUP Grading System -- Epstein et al. 2016