Fully Automatic Left Atrium Segmentation From Late Gadolinium Enhanced Magnetic Resonance Imaging Using a Dual Fully...

Area of Science:

• Cardiology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Atrial fibrillation (AF) is the most common cardiac arrhythmia, with current treatments limited by incomplete understanding of atrial structures.
• Manual segmentation of 3D cardiac MRI (LGE-MRI) for AF analysis is labor-intensive and error-prone, hindering research and clinical application.

Purpose of the Study:

• To develop a robust, automated method for segmenting left atrial (LA) structures from 3D LGE-MRIs.
• To improve the accuracy and efficiency of atrial analysis for better understanding and treatment of AF.

Main Methods:

• Development of AtriaNet, a 16-layer convolutional neural network (CNN) with a multi-scaled, dual-pathway architecture.
• Training and validation on 154 3D LGE-MRIs from AF patients, utilizing computationally efficient batch prediction for rapid processing.

Main Results:

• AtriaNet achieved high segmentation accuracy with DICE scores of 0.940 (epicardium) and 0.942 (endocardium).
• The model demonstrated minimal inter-patient variance (<0.001) and processed each MRI in under 1 minute.
• Accurate estimations of LA diameter (within 1.59 mm) and volume (within 4.01 cm³).

Conclusions:

• AtriaNet represents the most robust approach to date for segmenting LGE-MRIs, outperforming state-of-the-art CNNs.
• Automated segmentation significantly enhances atrial reconstruction and analysis, potentially leading to improved understanding and treatment strategies for AF.