New international research affirms the value of adjunctive artificial intelligence (AI) in improving prostate cancer (PCa) detection on biparametric magnetic resonance imaging (bpMRI) in contrast to unassisted radiologist interpretation.
For the retrospective study, recently published in the American Journal of Roentgenology, researchers compared adjunctive use of a deep learning PCa detection algorithm (developed by the Molecular Imaging Branch of the National Cancer Institute) to unassisted radiologist interpretation of bpMRI scans. The cohort was comprised of 180 patients (including 60 control patients) who had initially had multiparametric MRI (mpMRI) exams and prostate biopsy or radical prostatectomy, according to the study.
At a threshold of PI-RADS > 3, the study authors found that adjunctive AI had a 77.2 percent lesion-level positive predictive value (PPV) for clinically significant prostate cancer (csPCa) in comparison to 67.2 percent for unassisted interpretation. The researchers noted an 11.5 percent increase in adjunctive AI lesion-level PPV for PCa (80.9 percent) in contrast to radiologist interpretation (69.4 percent) at the same threshold.
Here one can see MRI images, AI segmentation mapping and histopathology annotations revealing a region of Gleason 4 + 3 prostate cancer in a 77-year-old man with a serum prostate-specific antigen (PSA) level of 17.8 ng/mL. (Images courtesy of the American Journal of Roentgenology,)
However, adjunctive AI had a lower lesion-level sensitivity rate than unassisted radiologist interpretation for both csPCa (44.4 percent vs. 48 percent) and PCa (41.7 percent vs. 44.9 percent) at the PI-RADS > 3 threshold.
“Our study found that AI assistance resulted in significantly improved lesion-level PPV for detecting both csPCa and PCa compared with unassisted reads. However, lesion-level sensitivity was slightly lower with AI assistance, across all readers. This suggests that although AI had issues in detecting some cancerous lesions using radical prostatectomy pathology as the reference standard, lesions identified by AI were more likely to be true-positives,” wrote study co-author Baris Turkbey, M.D., a senior clinician/radiologist at the National Cancer Institute and National Institutes of Health (NIH), and colleagues.
Three Key Takeaways
- Adjunctive AI significantly improves lesion-level PPV. AI-assisted interpretation of bpMRI raised lesion-level positive predictive value (PPV) for clinically significant prostate cancer (csPCa) from 67.2 percent (radiologist alone) to 77.2 percent, and for overall PCa from 69.4 percent to 80.9 percent.
- Sensitivity remains slightly lower with adjunctive AI. Despite higher PPV, AI-assisted reads had a marginally lower lesion-level sensitivity compared to radiologist-only interpretation for both csPCa (44.4 percent vs. 48 percent) and overall PCa (41.7 percent vs. 44.9 percent).
- AI enhances inter-reader agreement. Adjunctive AI substantially improved consistency among radiologists, increasing inter-reader agreement on lesion-level PI-RADS scores from 33.6 percent to 74.8 percent and patient-level scores from 50.7 percent to 70.4 percent.
The study authors pointed out that adjunctive AI significantly enhanced inter-reader agreement for lesion-level PI-RADS scores (74.8 percent vs. 33.6 percent) and patient-level PI-RADS scores (70.4 percent vs. 50.7 percent) in contrast to unassisted reading by radiologists.
“These findings contribute to the ongoing evaluation of AI in clinical practice, supporting its role as a valuable tool for reducing variability among readers and improving reliability of PCa diagnosis on MRI,” noted Turkbey and colleagues.
(Editor’s note: For related content, see “Assessing the Impact of Adjunctive and Stand-Alone AI for Prostate MRI,” “Multinational Study Reaffirms Value of Adjunctive AI for Prostate MRI” and “Study: AI-Generated ADC Maps from MRI More Than Double Specificity in Prostate Cancer Detection.”)
In regard to study limitations, the authors acknowledged the possibility of missed cancers in some of the control patients, retrospective processing of bpMRI images through exclusion of dynamic contrast-enhanced (DCE)images from original mpMRI exams, and all reviewing radiologists being affiliated with large academic hospital facilities.
