Automatic Explainable Segmentation of Abdominal Aortic Aneurysm From Computed Tomography Angiography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Abdominal aortic aneurysm (AAA) screening and analysis require accurate segmentation of vascular structures.
• Current segmentation methods can be time-consuming and may lack automation.
• Deep learning offers potential for automated and efficient medical image analysis.

Purpose of the Study:

• To develop and validate an automated deep learning framework for segmenting abdominal aortic aneurysms (AAA) in contrast-enhanced computed tomography angiography (CTA) images.
• To assess the framework's accuracy, efficiency, and potential for clinical application in AAA screening and analysis.

Main Methods:

• Development of a deep learning framework utilizing a dynamic router and three specialized U-Net models for AAA segmentation.
• Training and validation on a large dataset (9,080 images) and testing on a separate set (1,560 images).
• Evaluation of segmentation performance using Dice scores (DS), Intersection over Union (IoU), and Hausdorff distance (HD95).
• Assessment of processing time for automated segmentation and manual refinement using non-uniform rational B-splines (NURBS).

Main Results:

• The framework achieved high segmentation accuracy for both aortic lumen and outer wall, with DS of 0.9648/0.9615 and IoU of 0.9324/0.9264.
• Excellent performance was indicated by HD95 values of 1.3490 mm and 1.3670 mm for lumen and wall, respectively.
• The automated system processed images rapidly at approximately 17 ± 1 milliseconds per frame.
• Manual NURBS refinement for complex cases ranged from 3 to 20 seconds per frame.

Conclusions:

• The automated deep learning framework demonstrates high accuracy and efficiency for segmenting abdominal aortic aneurysms in CTA images.
• The system's real-time processing capability makes it suitable for clinical screening and analysis.
• Further research integrating multimodal imaging and optimizing NURBS refinement could enhance accuracy and efficiency.