Emerging prostate MRI research suggests that a deep learning model may provide utility in ruling out cases of clinically significant prostate cancer (csPCa).
For the retrospective study, recently published in Insights into Imaging, researchers compared the use of the deep learning model (MR-Prostate AI research application v1.3.4, Siemens Healthineers) versus a reference standard of original radiologist reports for 442 adult men (mean age of 65) with suspected prostate cancer (PCa). The deep learning model was trained on 3,087 prostate biMRI exams, according to the study.
For PI-RADS > 4 lesions, the researchers found that the deep learning model offered a 81.2 percent specificity rate and an 82.9 percent negative predictive value (NPV). For PI-RADS > 5 lesions, the deep learning model provided a 93.8 percent specificity and a 97.3 percent NPV.
“The high NPV of the DL-CAD for ruling out PI-RADS 4–5 lesions shows that patients classified as negative by the algorithm have a low risk of harboring clinically significant PCa, which could reduce unnecessary biopsies and their associated risks in a clinical setting,” noted lead study author Deepak Jain, MD, who is affiliated with the Department of Diagnostic, Molecular and Interventional Radiology at the Icahn School of Medicine in Mount Sinai, N.Y., and colleagues.
However, the study authors noted that the deep learning model demonstrated a fair AUC (67 percent) for per-patient detection of PI-RADS > 3 lesions. They also pointed out per-patient sensitivity and positive predictive value (PPV) of 65.3 percent and 62.7 percent, respectively, for PI-RADS > 4 lesions, and a 65.7 percent PPV for identifying PI-RADS > 5 lesions.
Three Key Takeaways
• Strong rule-out capability for csPCa: The deep learning model demonstrated high specificity and negative predictive value (NPV) — up to 97.3 percent for PI-RADS ≥5 — supporting its potential role in safely reducing unnecessary prostate biopsies in patients classified as having negative prostate MRI findings.
• Limited standalone detection performance. Despite strong rule-out performance, the model showed only fair overall discrimination (AUC 67 percent) and modest sensitivity (~65 percent), indicating it should complement—not replace—radiologist interpretation, particularly for detecting PI-RADS 3 and PI-RADS 4 lesions.
• Performance gaps tied to lesion characteristics. Most false positives were due to benign conditions (e.g., BPH, scarring) while false negatives were more common in challenging anatomical regions (apex, PZ/TZ junction), highlighting the need for improved training datasets and caution with lesion localization and staging applications.
An analysis of the discordance between the deep learning model and radiologist reports found that benign prostatic hyperplasia (BPH), PI-RADS 2 presentations and scarring accounted for 83 percent of false positive results with the deep learning model. The researchers also noted that lesion location played a role with deep learning false negatives as 28 percent of the cases involved lesions in the prostate apex, peripheral zone (PZ) or the PZ/transitional zone (TZ) junction.
“.. The relatively low sensitivities observed suggest that the DL-CAD could benefit from additional training on more cases with BPH nodules, lesions located in the PZ-TZ junction, apex, and sub-cm lesions,” added Jain and colleagues. “This is especially relevant for lesion-based detection and classification, which could be potentially useful in disease staging and targeted biopsies, where accurate lesion localization is essential.”
(Editor’s note: For related content, see “Can AI Bolster Prostate Biopsy Efficiency?: What a Multicenter MRI Study Reveals,” “Meta-Analysis Shows Prognostic Impact of Prostate MRI Prior to Radical Prostatectomy” and “5T Prostate MRI Study Reveals Enhanced Image Quality and Detection.”)
Beyond the inherent limitations of a single-center retrospective analysis, the authors conceded the lack of histopathologic correlation and the use of scanners from only two manufacturers.
