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

Figures

• 

  Figure 1. Flowchart of the study population inclusion and exclusion process.

• 

  Figure 2. Schematic overview of the proposed two-stage deep learning framework. The pipeline consists of two sequential stages for automated pulmonary embolism detection. Stage 1 (Candidate Generation) utilizes a DuckNet architecture to process 2D slices from the resampled CT volume, generating initial binary masks with high sensitivity. Stage 2 (False-Positive Reduction) extracts 3D volumetric patches corresponding to the candidate masks.

• 

  Figure 3. Illustration of the two-stage Modified Mask R-CNN framework. Stage 1 generates initial candidates (red and yellow arrows) with high sensitivity. Stage 2 employs a False-Positive Reduction (FPR) module to analyze 3D morphological features, effectively eliminating false positives (red arrows) while retaining true emboli (yellow arrows).

• 

  Figure 4. Comparative performance on the internal validation dataset. The Modified Mask R-CNN (orange) achieved a substantial reduction in false-positive rates per scan compared to the RPN-only model (blue), while maintaining high patient-level sensitivity.

• 

  Figure 5. Representative examples of false-positive reduction by the Modified Mask R-CNN. Three common sources of false positives in automated pulmonary embolism detection are shown: hilar lymph nodes mimicking filling defects (top row), contrast-enhanced pulmonary veins (middle row), and beam-hardening artifacts near the superior vena cava (bottom row). Left column: Original axial CTPA images. Middle column: RPN-only model output with false-positive detections indicated by the red circles. Right column: Modified Mask R-CNN output demonstrating the successful elimination of false positives using the dedicated FPR module. The white arrows indicate the anatomical structures that were incorrectly identified by the RPN-only model but correctly filtered by the two-stage framework.

• 

  Figure 6. Segmentation accuracy comparison. Green contours represent expert radiologist annotations (Ground Truth), and red contours represent the model’s predictions. The Modified Mask R-CNN demonstrates high concordance with manual annotations, accurately delineating embolus boundaries even in complex anatomical locations like the main pulmonary artery bifurcation.