Left Atrial Appendage Segmentation Using Fully Convolutional Neural Networks and Modified Three-Dimensional...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Thrombosis is a global health concern, with the left atrial appendage (LAA) being a primary source of thrombi in atrial fibrillation (AF) patients.
• LAA morphology and volume correlate with AF risk and thromboembolic events, highlighting the need for accurate LAA assessment.
• Automated LAA segmentation is crucial for efficient diagnosis and risk stratification in clinical practice.

Purpose of the Study:

• To develop a robust method for automatic segmentation of the left atrial appendage (LAA) from computed tomographic angiography (CTA) data.
• To address the challenges posed by significant anatomical variations in the LAA.
• To create a computationally efficient tool for routine clinical use.

Main Methods:

• Utilized fully convolutional neural networks (FCNs) adapted from natural image segmentation for 2-D LAA slice segmentation.
• Employed a modified dense 3-D conditional random field (CRF) to refine segmentations by incorporating 3-D spatial and contextual information.
• Fine-tuned pre-trained FCN models on CTA data after manual region of interest (ROI) localization.

Main Results:

• Achieved a mean Dice overlap of [insert value] and a mean volume overlap of [insert value] compared to manual annotations.
• Demonstrated high robustness in segmenting the LAA despite considerable anatomical variations.
• Exhibited computational efficiency with a processing time of less than 40 seconds per CTA volume.

Conclusions:

• The proposed FCN and 3-D CRF method provides accurate and robust automatic LAA segmentation on CTA data.
• The method's efficiency and accuracy make it suitable for integration into daily clinical workflows for AF diagnosis and management.
• This automated approach facilitates better prediction of thromboembolic risk in AF patients.