Atrial fibrillation (AF), a common cardiac arrhythmia affecting a significant proportion of the population, is a leading cause of stroke and thromboembolism. The left atrium (LA) and its appendage (LAA) are critical sites for thrombus formation. Computational fluid simulations using LA geometries from patient-specific images have added value in understanding blood flow patterns, and thrombogenesis risk in the LA. Usually, the required 3D anatomical reconstructions of the LA are obtained from segmentations performed on computed tomography (CT) and magnetic resonance imaging (MRI). However, these imaging modalities are not available for all AF patients at risk of stroke. Ultrasound (US) imaging offers a more affordable alternative, yet its application for reconstructing 3D anatomical structures, particularly the LA and LAA, remains a challenge. This study investigates the feasibility of reconstructing the LA and LAA in 3D from sparse 2D planes using advanced deep learning methods. Two reconstruction pipelines were evaluated: E-Pix2Vox++, trained with sparse 2D TEE slices; and neural implicit functions with shape prior, trained with sparse 2D slices from axial, coronal, and sagittal planes. Both models were trained on synthetic datasets generated from a statistical shape model of the LA built from segmentations of 128 patient-specific scans, comprising an augmented dataset of over 1200 synthetic meshes. The E-Pix2Vox++ and the neural implicit functions algorithms (based only on four 2D planes) achieved an average intersection over union (IoU) score of 0.748 and 0.789, respectively, while the average Hausdorff Distance was of 7.091 mm and 1.645 mm, respectively, with main differences arising around the pulmonary veins and the tip of the LAA. This work highlights the promise of leveraging deep learning to reconstruct 3D LA geometries from ultrasound data, which will contribute to democratize the use computational fluid simulations to support clinical decisions of AF patients in most hospitals.