Convolutional Neural Network based Segmentation of Abdominal Aortic Aneurysms.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Surgery

Background:

• Abdominal aortic aneurysms (AAAs) are life-threatening dilations of the aorta with high mortality rates.
• Accurate identification and characterization of AAAs are crucial for predicting rupture risk and guiding interventions.
• Current methods for AAA segmentation can be time-consuming and may lack precision.

Purpose of the Study:

• To investigate the feasibility of automating the segmentation of abdominal aortic aneurysms (AAAs) and related structures from computed tomography angiograms (CTAs).
• To develop and evaluate a 3D U-Net convolutional neural network (CNN) model for precise AAA detection and segmentation.
• To assess the impact of training data size and transfer learning on model accuracy and consistency.

Main Methods:

• Utilized 30 patient-specific CTAs with corresponding binary masks for training and testing a 3D U-Net CNN model.
• Employed a Dice Similarity Coefficient (DSC) based loss function to measure segmentation accuracy.
• Determined optimum probability thresholds (OPTs) for voxel-level outputs and used 3D volume rendering for validation.

Main Results:

• The trained 3D U-Net models achieved visually accurate automatic segmentations of AAAs, aneurysm sacs, thrombus, and calcifications.
• Model convergence and segmentation accuracy improved with increased training sample size.
• Transfer learning enhanced DSC loss, with median OPTs approaching 0.5 as training data increased.

Conclusions:

• Automated segmentation of AAAs using a 3D U-Net CNN is feasible and accurate.
• Increasing training data and employing transfer learning significantly improve model performance.
• This AI-driven approach holds promise for enhancing clinical risk assessment and timely intervention for AAAs.