Segmentation of the mitral valve (MV) across the cardiac cycle is pivotal for a comprehensive assessment of its functionality and for initialization of numerical models. Achieving a full assessment of MV dynamics through 4D transesopahgeal echocardiography (4DTEE) involve challenges which may be addressed with deep learning techniques. In this work, we propose an innovative usage of positional encoding to achieve full-cycle MV segmentation from 4DTEE using a convolutional neural network (CNN) supervised with only the end-systole and end-diastole frames. Anterior and posterior leaflets were manually annotated from 120 intraoperative 4DTEE acquisitions. A modified 3D residual UNet was trained on this dataset, integrating temporal dynamics by encoding each frame's position within the cycle. The proposed method achieved an average Dice score of 0.79 for end-systole and end-diastole, with surface and Hausdorff distances of 0.98 mm and 4.43 mm, respectively. Intermediate frames had slightly lower performance (Dice Score: 0.77, surface and Hausdorff distances: 0.99 mm and 5.56 mm). However, it significantly outperformed a model without temporal encoding. These results, in line with prior studies, show the model effectively recognizing MV structures across the cardiac cycle, despite slightly reduced accuracy in intermediate frames. This study pioneers comprehensive MV dynamics tracking from 4DTEE data using positional encodings. Future work aims to improve the model by integrating unlabeled frames into the training process