3D Reconstruction of the Left Atrial Appendage from Multiple 2D Echocardiographic Data

Abstract

Clot formation in the left atrial appendage (LAA) is a very common problem in patients with previous pathologies, such as atrial fibrillation (AF), which can lead to fatal consequences. In these patients, computer simulations can help predicting possible risks and assessing clinicians’ decisions. However, to perform LAA fluid simulations, volumetric images are necessary, from which the 3D structure can be segmented. Computed tomography (CT) scans are useful for this, but due to their high cost, high dose of radiation or simply because they are not always available in every hospital, they are performed in a minimum number of patients. On the other hand, echocardiography is a common technique, which is done routinely in AF patients. The objective of this project is to use deep learning (DL) techniques to reconstruct a 3D model of the left atrial appendage from single multi-view echocardiography images. The hypothesis to be tested is that a correctly trained DL algorithm can learn to reconstruct a 3D structure from ultrasound images, while keeping the most important details from each of the input views, thus obtaining a realistic enough model. A modification of classic variational autoencoders was considered an optimal approach for the task of 3D reconstruction in echocardiography images. The performance of the network was evaluated by computing the Dice Score and Hausdorff distance, and comparing fluid simulation results between ground-truth and reconstructions. The results demonstrate the potential of DL and variational autoencoders in reconstruction problems and open the possibility of cheaper and more efficient stratification method for patients with AF.