3D Pulmonary Artery Segmentation from CTA Scans Using Deep Learning with Realistic Data Augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Accurate characterization of mediastinal vasculature, particularly the pulmonary artery (PA), is crucial for diagnosing pulmonary vascular diseases.
• Automated segmentation of the PA in computed tomography angiography (CTA) images is needed for advanced geometrical analysis and clinical applications.

Purpose of the Study:

• To develop and validate a novel 3D convolutional neural network (CNN) for automatic segmentation of the pulmonary artery from CTA images.
• To enable detailed analysis of PA geometry in both healthy and diseased states, supporting fluid mechanics models and personalized medicine.

Main Methods:

• A new 3D CNN architecture was designed and trained on diverse patient cohorts.
• A robust data augmentation strategy using principal component analysis on deformation fields from affine registration was employed.
• The network's performance was validated on 91 datasets against semi-automatically delineated ground truths.

Main Results:

• The proposed 3D CNN achieved high accuracy in PA segmentation.
• Quantitative evaluation yielded mean Dice coefficient of 0.89, Jaccard coefficient of 0.80, and mean surface distance of 1.25 mm.
• Performance was benchmarked against a standard Unet architecture, demonstrating competitive or superior results.

Conclusions:

• The developed 3D CNN provides an effective tool for automated pulmonary artery segmentation in CTA.
• This automated approach facilitates quantitative analysis of PA morphology for clinical research and patient-specific modeling.
• The study highlights the potential of deep learning in advancing cardiovascular imaging analysis.