Deep learning-based cardiac computed tomography angiography left atrial segmentation and quantification in atrial...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Quantitative assessment of left atrial volume (LAV) is crucial for understanding atrial fibrillation pathogenesis.
• Automated left atrial segmentation and LAV measurement present significant challenges in clinical practice.

Purpose of the Study:

• To identify the optimal deep learning model for left atrial segmentation using cardiac computed tomography angiography (CTA).
• To perform accurate quantitative LAV measurements using the best-performing segmentation model.

Main Methods:

• A multi-center cohort of 182 cardiac CTAs from patients with atrial fibrillation was curated.
• Five state-of-the-art deep learning models (DAResUNet, nnFormer, xLSTM-UNet, UNETR, VNet) were trained and validated for left atrial segmentation.
• The optimal model was selected based on segmentation performance metrics and used for LAV consistency validation.

Main Results:

• DAResUNet demonstrated superior performance in Dice Similarity Coefficient (DSC) and Jaccard Index (JI).
• VNet excelled in Hausdorff Distance (HD) and Average Surface Distance (ASD) metrics.
• Bland-Altman analysis confirmed strong agreement between automated and manual LAV measurements (mean bias -5.69 mL).

Conclusions:

• Deep learning models, trained on a cohort of 182 CTA images, achieve competitive results in left atrial segmentation.
• Automated LAV assessment using deep learning shows potential as a biomarker for predicting atrial fibrillation onset.