3D region-growing nnU-Net improves pulmonary embolism detection on CTPA: a dual-cohort validation study
Revolutionizing PE Detection: A 3D AI Approach for Enhanced Accuracy
Leveraging advanced nnU-Net models with region-growing for superior pulmonary embolism identification and quantification on CTPA.
Executive Impact: Pioneering AI in Radiology
This study rigorously compared three nnU-Net models for automated pulmonary embolism (PE) detection and blood clot volume (BCV) quantification on CTPA. The 3D nnU-Net integrating region-growing (Model C) emerged as the superior performer, achieving significantly higher AUROC scores compared to 2D models on both internal (0.909 vs. 0.784-0.816) and external (0.868 vs. 0.843-0.846) datasets. While BCV was not found to predict short-term clinical outcomes like MACE or survival, the enhanced model's ability to precisely delineate emboli provides a robust foundation for future personalized risk stratification and improved patient outcomes, aligning with the need for data sovereignty and tailored AI solutions.
0.000
Internal AUROC (3D Model C)
0.000
External AUROC (3D Model C)
0
External Sensitivity (0mm³ cutoff)
0
External Accuracy (0mm³ cutoff)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Model Performance
Methodology
Comparative Analysis
Clinical Relevance
The 3D nnU-Net with region-growing (Model C) consistently outperformed 2D models in accuracy and generalizability for PE detection. It achieved an AUROC of 0.909 internally and 0.868 externally, demonstrating its robust capability. While BCV was not predictive of MACE or survival, the model's high sensitivity (83.6%) and accuracy (79.5%) on external data at a zero-clot threshold make it highly suitable for high-sensitivity screening in emergency settings.
0.909
Model C Internal AUROC
The study utilized a dual-cohort validation strategy, training models on a local Humanitas dataset and independently testing on the public RSPECT-RSNA dataset. Three distinct nnU-Net models were compared: baseline 2D (Model A), 2D with region-growing (Model B), and 3D with region-growing (Model C). Region-growing with a 5-voxel negative buffer was crucial for improving segmentation consistency and differentiating thrombi from adjacent tissues.
Enterprise Process Flow
CTPA Archive Screening (9,715 scans)
→
Pre-screen with legacy PE model (BCV > 800 µL)
→
Expert Review & Manual Segmentation
→
nnU-Net Model Training (2D/3D + Region-Growing)
→
Internal & External Testing (Humanitas & RSPECT-RSNA)
A detailed comparison of Model C against 2D models (A and B) revealed superior performance. The 3D approach's ability to process spatial information across slices, coupled with region-growing, effectively addressed challenges like confounding structures and motion artifacts. This table summarizes key performance metrics across models and datasets.
| Model | Internal AUROC | External AUROC | External Sensitivity (0mm³) | External Accuracy (0mm³) |
|---|---|---|---|---|
| Model A (2D Baseline) | 0.784 | 0.843 | 0.836 | 0.795 |
| Model B (2D + RG) | 0.816 | 0.846 | 0.920 | 0.592 |
| Model C (3D + RG) | 0.909 | 0.868 | 0.836 | 0.795 |
While the 3D nnU-Net significantly improves PE detection, the study found no statistically significant correlation between automated BCV and MACE or survival outcomes. This suggests that BCV in isolation may be an insufficient biomarker for risk stratification. Future models should integrate BCV with other imaging biomarkers (e.g., right-to-left ventricular diameter ratio) and clinical data to build more comprehensive risk stratification tools.
Optimizing Clinical Impact: Beyond Volume
The current study's findings on the lack of correlation between blood clot volume (BCV) and major adverse cardiovascular events (MACE) or survival time underscore a critical insight for enterprise AI deployment in healthcare. While the AI model excels at detecting and quantifying PE, its immediate clinical utility for risk stratification based solely on BCV is limited. This means that to achieve true clinical impact, AI solutions must move beyond simple quantification and integrate a broader range of patient-specific data.
Challenge:
Develop AI solutions that combine image-derived biomarkers (like BCV, ventricular strain) with electronic health record data (comorbidities, patient history) for a holistic, personalized risk assessment.
Solution:
Our 3D nnU-Net provides the highly accurate, foundational PE detection. The next step involves leveraging this foundation to build a multi-modal AI system that correlates detailed imaging features with clinical parameters. This integrated approach will enable more precise patient prognoses and tailored treatment strategies, ultimately leading to improved patient outcomes and more efficient resource allocation within healthcare systems. This demonstrates the necessity for adaptable AI frameworks that can evolve with clinical understanding.
Calculate Your Potential ROI
Estimate the efficiency gains and cost savings AI can bring to your operations. Adjust the parameters to see a customized impact for your enterprise.
Annual Savings
$0
Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A structured approach to integrating cutting-edge AI, from initial data integration to long-term monitoring and optimization.
Phase 1: Local Data Integration & Refinement
Establish secure, compliant pipelines for integrating institutional CTPA data. Fine-tune initial 3D nnU-Net models using local expert-annotated datasets, focusing on regional variations and scanner specifics to maximize accuracy and clinical relevance.
Phase 2: Multi-modal AI Development
Expand beyond BCV. Integrate automated extraction of other critical biomarkers (e.g., right ventricular strain, pulmonary artery diameter) and link with de-identified patient EHR data (comorbidities, MACE history) to build a comprehensive risk stratification model.
Phase 3: Prospective Validation & CE Marking Compliance
Conduct prospective multicenter validation studies to confirm real-world performance and generalizability. Work towards European conformity (CE marking) to ensure regulatory compliance and facilitate broader clinical adoption.
Phase 4: Clinical Workflow Integration & Monitoring
Seamlessly integrate the validated AI solution into existing radiology workflows. Implement continuous monitoring and feedback mechanisms to track performance, identify biases, and ensure ongoing optimization and adaptation to evolving clinical needs.
Ready to Transform Your Enterprise with AI?
Book a personalized strategy session with our AI experts to explore how these insights can be tailored to your organization's unique needs and objectives.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat