Emerging findings from a multinational study demonstrate increased sensitivity, specificity and the area under the receiver operating characteristic curve (AUC) with adjunctive artificial intelligence (AI) for clinically significant prostate cancer (csPCa) detection on biparametric MRI (bpMRI) prostate exams.
For the retrospective study, recently published in JAMA Network Open, researchers evaluated the use of AI software for csPCa detection in 360 men with PI-RADS 3 or higher presentations on prostate MRI. The study authors noted that 122 men in the cohort were diagnosed with csPCa. Drawing the cohort from 53 facilities in 17 countries, the researchers noted that the AI software, developed within the Prostate Imaging-Cancer AI (PI-CAI) Consortium, was utilized by 61 radiologists, including 27 non-experts and 34 readers who had interpreted more than 1,000 prostate MRI cases in their career and over 200 cases a year.
The study authors found that adjunctive AI increased sensitivity by 2.5 percent (96.8 percent vs. 94.3 percent) and specificity by 3.4 percent (50.1 percent vs. 46.7 percent) in contrast to unassisted radiologist interpretation of bpMRI.
Adjunctive AI for prostate bpMRI also demonstrated a 91.6 percent AUC in comparison to 88.2 percent for unassisted AI, and reduced false positive cases by 10, according to the researchers.
“With AI assistance, PI-RADS scores were updated in 33% of assessments, including 17% involving reclassification between positive and negative MRI results, which likely altered the biopsy decision for these patients. AI assistance was associated with improved csPCa detection in PI-RADS categories 4 and 5 and reduced detection in the PI-RADS category 1 to 2 from 6% to 3%,” wrote lead study author Jasper J. Twilt, MSc, who is affiliated with the Minimally Invasive Image-Guided Intervention Center and the Department of Medical Imaging at the Radboud University Medical Center in Nijmegen, the Netherlands, and colleagues.
Three Key Takeaways
1. Improved diagnostic performance. Adjunctive AI use on biparametric MRI increased sensitivity (by 2.5 percent) and specificity (by 3.4 percent) for detecting clinically significant prostate cancer (csPCa), along with a higher AUC (91.6 percent vs. 88.2 percent), indicating better overall diagnostic accuracy.
2. Impact on PI-RADS scoring and biopsy decisions. AI-assisted interpretation led to PI-RADS score changes in 33 percent of cases with 17 percent involving reclassification between positive and negative MRI findings, potentially influencing biopsy decisions and improving csPCa detection in PI-RADS 4–5, while reducing false positives in PI-RADS 1–2.
3. Greater benefit for less experienced radiologists. Non-expert radiologists showed larger gains from AI support, achieving higher sensitivity and specificity improvements than experts. Notably, non-experts using AI outperformed experts without AI, suggesting AI may help reduce variability in diagnostic performance.
The researchers also found that non-expert radiologists had higher sensitivity and specificity gains with adjunctive AI in comparison to expert radiologists. There was a 3.7 percent increase in sensitivity for non-experts in contrast to 1.5 percent for expert radiologists, according to the study authors. They noted a 4.3 percent boost in specificity for non-experts versus 2.8 percent for expert readers.
“ … Our study suggests that non-experts experienced a greater performance boost from AI assistance compared with experts, highlighting the potential of AI to reduce performance differences between experts and nonexperts. Non-experts with AI support achieved higher AUROC scores than experts without AI, and their sensitivity surpassed that of experts in both unassisted and AI-assisted settings,” added Twilt and colleagues.
(Editor’s note: For related content, see “Study: AI-Generated ADC Maps from MRI More than Double Specificity in Prostate Cancer Detection,” “Study: MRI-Based AI Enhances Detection of Seminal Vesicle Invasion in Prostate Cancer” and “New bpMRI Study Suggests AI Offers Comparable Results to Radiologists for PCa Detection.”)
In regard to study limitations, the authors noted the retrospective nature of the research and conceded that the AI software was assessed with MRI scans largely performed with one MRI system. The researchers acknowledged the use of a controlled online reading workstation and noted a lack of assessment of the AI software with respect to clinical applicability and its potential impact on workflow efficiency.
