Prostate cancer remains a significant health concern, with early and accurate detection crucial for improving patient outcomes. Magnetic resonance imaging (MRI) has become a valuable tool in identifying clinically significant prostate cancer (csPCa), particularly through the Prostate Imaging-Reporting and Data Systems (PI-RADS) scoring system. However, detection performance varies across centres due to differences in radiologist expertise, imaging protocols and scanner technology.
Artificial intelligence-based deep-learning computer-aided detection (DL-CAD) systems offer a promising solution to improve diagnostic consistency. A recent study published in European Radiology evaluates a CE-marked AI software designed to detect csPCa, comparing its performance to multidisciplinary team (MDT)-supported radiologists across multiple hospitals, scanner models and vendors.
Multi-Centre Study and AI Model Validation
The study assessed an AI-powered DL-CAD system for csPCa detection using multiparametric MRI (mpMRI). The AI model, trained on data from multiple UK hospitals and the PROSTATEx dataset, was validated separately across six hospital sites with different MRI scanners. The validation set included 252 patients scanned between 2018 and 2022, with a csPCa prevalence of 31%. Biopsy results served as the reference standard, and patients with negative MRI and low prostate-specific antigen density (PSAD) were assumed cancer-free. The study used receiver operating characteristic (ROC) analysis to compare AI performance with radiologists using PI-RADS/Likert scoring.
Recommended Read: Optimising Prostate Cancer Diagnosis with MRI-Based Risk Stratification
The AI model achieved an area under the curve (AUC) of 0.91, closely matching radiologists’ AUC of 0.95. At the predetermined risk threshold of 3.5, AI demonstrated a sensitivity of 95% and specificity of 67%, while radiologists achieved 99% sensitivity and 73% specificity. The AI system performed consistently across different scanner models and field strengths, confirming its generalisability.
Performance Across Sites and Scanner Variability
A critical aspect of this validation study was the AI’s ability to generalise across multiple hospital settings. Performance was evaluated at both patient and lesion levels, with site-specific ROC analysis revealing AUC values of at least 0.83 at all locations. Notably, at the fully held-out validation site, the AI model achieved an AUC of 0.92, reinforcing its robustness. However, sensitivity and specificity varied across sites, suggesting that site-specific calibration might enhance AI reliability in diverse clinical environments.
At the lesion level, AI missed 14% of GG≥2 lesions, compared to 7% missed by radiologists. Additionally, AI identified more false-positive lesions, which could lead to unnecessary biopsies. While these results indicate AI’s potential as a decision-support tool, they also highlight the need for careful integration into clinical workflows, ensuring radiologists oversee its outputs to minimise diagnostic errors.
Implications and Future Directions
The findings suggest that AI-assisted prostate cancer detection could support radiologists in identifying csPCa with high accuracy, potentially reducing variability in diagnostic outcomes. AI’s ability to integrate into existing radiology workflows without requiring significant changes enhances its feasibility for widespread adoption. However, further prospective studies are needed to assess AI’s impact on clinical decision-making, particularly in guiding biopsy procedures based on AI-detected lesions.
Future research should explore AI’s performance across broader patient populations, different ethnicities and varying disease prevalence levels. Additionally, investigating AI’s role in active surveillance and long-term patient outcomes will be essential for determining its full clinical utility. Randomised controlled trials comparing AI-driven biopsy strategies with standard radiologist-driven approaches will provide deeper insights into AI’s potential benefits and limitations.
This multi-centre validation study demonstrates that AI-powered DL-CAD for prostate MRI achieves non-inferior diagnostic performance compared to MDT-supported radiologists. The AI model generalises well across multiple hospitals and scanner types, offering a promising tool for standardising prostate cancer detection. While AI can enhance diagnostic consistency, further studies are necessary to refine its application and evaluate its long-term clinical impact.
Source: European Radiology
Image Credit: iStock
