Simulating workload reduction with an AI-based prostate cancer detection pathway using a prediction uncertainty...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Prostate cancer diagnostics

Background:

• Prostate cancer detection relies heavily on radiologist interpretation of MRI scans.
• Reducing radiologist workload while maintaining diagnostic accuracy is a key challenge.
• Uncertainty quantification (UQ) metrics can help identify AI predictions requiring human review.

Purpose of the Study:

• To compare two UQ metrics (meanUQ and varUQ) for ruling out prostate MRI scans with high-confidence AI predictions.
• To assess the potential reduction in radiologist workload in a clinically significant prostate cancer (csPCa) detection pathway.
• To investigate the efficacy and accuracy of an AI-based rule-out pathway compared to standard radiologist reading.

Main Methods:

• Retrospective analysis of 1612 prostate MRI scans from three institutes.
• Comparison of standard diagnostic pathway versus an AI-based rule-out pathway using 15 AI submodels.
• Evaluation of meanUQ and varUQ prediction metrics using DeLong test on AUROC.
• Determination of workload reduction based on maintained sensitivity at non-inferior specificity margins (0.05 and 0.10).

Main Results:

• AI-based rule-out pathway demonstrated institute-specific workload reduction, up to 20% at a 0.10 non-inferiority margin.
• VarUQ showed higher, though non-significant, AUROC scores compared to meanUQ in specific cases.
• Non-significant workload reduction was observed at the 0.05 margin, highlighting the impact of the chosen threshold.

Conclusions:

• Both meanUQ and varUQ show potential for AI-based csPCa detection rule-out pathways.
• Utilizing varUQ in AI-driven pathways can decrease the number of scans requiring radiologist interpretation.
• Institute-specific calibration of UQ thresholds is crucial due to varying performance, enabling semi-autonomous AI assessment.