MRI-Based Deep Learning Model Bolsters Prediction of PI-RADS 3 and 4 Lesions

An emerging open-source deep learning model may provide enhanced risk stratification for the development of clinically significant prostate cancer (csPCa).

For a new retrospective study, recently published in European Radiology, the researchers reviewed prostate cancer (PCa) lesion probability maps generated by the deep learning model after assessment of biparametric magnetic resonance imaging (bpMRI) for 151 men (mean age of 65 and mean prostate-specific antigen (PSA) level of 8.3 ng/mL).¹

The study authors found that the deep learning model had an 86 percent AUC for predicting PI-RADS > 3 lesions and a 91 percent AUC for predicting PI-RADS > 4 lesions. The deep learning model detected 100 percent of PI-RADS > 4 lesions and 93 percent of PI-RADS > 3 lesions, according to the researchers.¹

“The AUC of 0.86 for PI-RADS ≥ 3 is noteworthy as it approaches the performance of models trained with PIRADS labels that use a threshold of ≥ 4, despite the inherent challenge of including PI-RADS 3 lesions, which are diagnostically challenging yet significant for biopsy consideration. … The model’s accuracy for predicting PI-RADS ≥ 4 is consistent with state-of-the-art models,” wrote lead study author Patricia M. Johnson, M.D., who is affiliated with the Bernard and Irene Schwartz Center for Biomedical Imaging and the Department of Radiology at the New York University Grossman School of Medicine in New York City, and colleagues.

Overall, the researchers noted a 78 percent AUC and an 87 percent sensitivity for detecting csPCa with the deep learning model. In comparison, three reviewing radiologists (with five, two and two years of experience with prostate MRI interpretation) provided sensitivity rates ranging from 57 to 84 percent) for PI-RADS > 3, according to the study authors.¹

However, the reviewing radiologists provided better specificity rates for csPCa detection (ranging between 67 to 87 percent) in contrast to 53 percent for the deep learning model.¹ In comparison to a 2024 study assessing MRI-based artificial intelligence (AI) for PCa detection, this study’s findings revealed higher specificity and lower sensitivity for the deep learning model, according to the researchers.²

“This suggests a more conservative approach to lesion classification, prioritizing specificity over sensitivity,” added Johnson and colleagues.

(Editor’s note: For related content, see “Adjunctive AI Bolsters Lesion-Level PPVs for csPCa in International bpMRI Study,” “Study: Adjunctive AI Provides Over 18 Percent Higher Lesion-Level Sensitivity on Prostate MRI” and “Multinational Study Reaffirms Value of Adjunctive AI for Prostate MRI.”)

In regard to study limitations, the authors acknowledged a short follow-up period for confirming negative MRI findings as well as a slightly higher T2W frequency-encoding resolution (.56 mm) than the resolution recommendation with PI-RADS v2.1 (.4 mm).¹

References

1. Johnson PM, Tong A, Ginocchio L, et al. External evaluation of an open-source deep learning model for prostate cancer detection on bi-parametric MRI. Eur Radiol. 2025 Aug 3. doi: 10.1007/s00330-025-11865-x. Online ahead of print.

2. Saha A, Bosma JS, Twilt JJ et al. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international,paired, non-inferiority, confirmatory study. Lancet Oncol. 2024;25(7):879-887.