Validation of UniverSeg for Interventional Abdominal Angiographic Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Vascular Disease Assessment

Background:

• Automatic segmentation of angiographic structures is crucial for vascular disease assessment.
• Current deep learning models lack validation on interventional angiographic data.

Purpose of the Study:

• To investigate the feasibility of angiographic segmentation using in-context learning with the UniverSeg model.
• To evaluate UniverSeg's performance on interventional fluoroscopic data without prior angiographic training.

Main Methods:

• Retrospective review of 234 patients undergoing celiac axis interventional fluoroscopy.
• Selection of 303 maximum contrast images for segmentation and binary mask creation.
• Testing UniverSeg with in-context learning in a fivefold nested cross-validation, analyzing performance based on arterial diameter and bifurcation number.

Main Results:

• Dice similarity coefficients ranged from 59.9% to 78.7% for decreasing arterial diameters.
• Balanced average Hausdorff distances were between 0.71 and 1.16 pixels.
• Performance improved with support class size, vessel diameter, and reduced bifurcations, plateauing at N=51.

Conclusions:

• UniverSeg is validated for arterial segmentation in interventional fluoroscopic procedures.
• The findings support the use of UniverSeg for vascular disease modeling and imaging research.
• In-context learning demonstrates the potential of cross-learning models for specialized medical imaging tasks.