Automatic initialization of the left atrial appendage centerline in 3D TEE images using deep learning

The Left Atrial Appendage (LAA) is a complex anatomical structure often responsible for cardiac thrombi. Transesophageal echocardiography (TEE) is the primary imaging modality for detailed LAA assessment and thrombus detection. Segmenting the LAA is crucial, yet existing 3D TEE-based methods often require manual centerline initialization, making them semi-automatic. This study aimed to automate the LAA centerline stage using deep learning (DL) techniques. A UNET-based DL model was trained and validated on 188 3D TEE volumes, with manual centerline annotations serving as the ground truth. Following automated centerline identification, state-of-the-art segmentation was applied, and the results were compared to manually corrected segmentations. The automated method achieved high segmentation performance, with an Average Distance Error of 0.60±0.46mm, a Dice Coefficient of 94.45±4.25%, and a 95% Hausdorff Distance of 2.21±1.86mm. These findings demonstrate the potential of DL methods to fully automate LAA segmentation in 3D TEE images.