A Promptable 3D-CT Foundation Model-Based Approach for Pulmonary Embolism

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Disease Research

Background:

• Blood clot volume (BCV) is a key biomarker for pulmonary embolism (PE) severity.
• Manual segmentation of thrombi on computed tomography angiography (CTA) is time-consuming, limiting clinical utility.
• Objective biomarkers are crucial for accurate PE assessment and treatment guidance.

Purpose of the Study:

• To evaluate ClotIA (Clot Interventional AI), a foundation model (FM)-based approach for rapid and interactive clot segmentation in PE.
• To assess the feasibility of AI-driven BCV quantification for clinical application.
• To overcome the limitations of manual segmentation in PE assessment.

Main Methods:

• A foundation model (RAPSv2) was fine-tuned on 309 PE patients from the RSPECT dataset.
• Segmentation performance was measured using the Dice similarity coefficient (DSC) and compared to nnUNet.
• The AI-predicted BCV was correlated with established imaging biomarkers of PE severity.

Main Results:

• ClotIA achieved a mean DSC of 0.83 after guided refinement, outperforming baseline and nnUNet.
• Predicted BCV demonstrated strong agreement with reference volumes (r=0.995).
• AI-quantified BCV significantly correlated with right ventricle/left ventricle (RV/LV) diameter (r=0.62) and volume ratios (r=0.68).

Conclusions:

• ClotIA facilitates rapid and reproducible 3D quantification of PE thrombi.
• The AI tool's findings correlate with established PE severity markers.
• ClotIA provides a foundation for translating novel biomarkers into clinical practice for PE management.