LA-Net: A Multi-Task Deep Network for the Segmentation of the Left Atrium

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Atrial fibrillation (AF) is a common arrhythmia with suboptimal treatment outcomes.
• Magnetic Resonance Imaging (MRI) offers valuable data for improving AF treatment, but lacks automated atrial segmentation tools.
• Accurate segmentation of the left atrium (LA) in cardiac MRI is crucial for patient management.

Purpose of the Study:

• To develop and evaluate LA-Net, a novel multi-task deep learning network for automated left atrial segmentation and edge mask generation from MRI.
• To improve the accuracy and efficiency of extracting LA information from cardiac MR images.

Main Methods:

• Proposed LA-Net, a multi-task network incorporating cross-attention modules (CAMs) and enhanced decoder modules (EDMs).
• Optimized LA-Net to simultaneously generate segmentation and edge masks from MRI data.
• Evaluated LA-Net on two MRI sequences: late gadolinium-enhanced (LGE) and balanced steady-state free precession (bSSFP).

Main Results:

• LA-Net achieved a Hausdorff distance of 12.43 mm and a Dice score of 0.92 on the LGE dataset.
• LA-Net achieved a Hausdorff distance of 17.41 mm and a Dice score of 0.90 on the bSSFP dataset.
• Performance surpassed existing methods like U-Net and SEGANet without post-processing.

Conclusions:

• LA-Net provides accurate and automatic left atrial segmentation from cardiac MRI.
• This tool can significantly aid in the management and treatment of patients with atrial fibrillation.
• Automated segmentation of cardiac structures from MRI holds promise for advancing cardiovascular diagnostics.