AI-Based Pulmonary Embolism Detection: The Added Value of a False-Positive Reduction Module over a Region Proposal...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Pulmonary Embolism Detection

Background:

• High false-positive rates challenge automated pulmonary embolism (PE) detection via CTPA.
• Evaluating a False-Positive Reduction (FPR) module integrated into a Region Proposal Network (RPN).

Purpose of the Study:

• Assess the added value of an FPR module in PE detection.
• Compare diagnostic performance of RPN-only vs. Modified Mask R-CNN with FPR.

Main Methods:

• Retrospective analysis of 303 CTPA scans (163 PE-positive, 140 PE-negative).
• External validation on 100 CTPA scans from RSNA PE Challenge dataset.
• Comparison of one-stage RPN-only model with two-stage Modified Mask R-CNN incorporating FPR.

Main Results:

• Modified Mask R-CNN significantly improved specificity, reducing false-positive rates by 31%.
• Positive Predictive Value increased by 10.5% with a slight reduction in sensitivity.
• Improved patient-level specificity for clinically significant emboli (≥ 1000 mm³).

Conclusions:

• Modified Mask R-CNN effectively reduces false positives in CTPA-based PE detection.
• The model maintains high sensitivity while enhancing diagnostic accuracy.
• This approach offers a valuable improvement over standard RPN models.