Cardiac anatomical segmentation is an essential step in many clinical applications that require quantitative measures such as myocardial mass and ventricle volume. Traditionally, these segmentations were manually performed by radiologists or clinical specialists, which is a time-consuming task and prone to inter-observer variability. In this context, deep learning (DL) appears as a useful emerging tool for this purpose, allowing for automated and semi-automated segmentations. However, DL models can still produce errors. Because of this, in some cases there is the need for post-processing methods to refine the output and improve the accuracy of the segmentation. This work proposes a denoising autoencoder (DAE) as a post-processing step to remove errors from the left atrium (LA) segmentations output by a U-Net model. The data used consisted of 82 ground truth LA masks segmented by clinical specialists from 3D DE-MRI scans from Hospital Clinic de Barcelona patients. For each ground truth, 10 to 20 synthetic erroneous masks were created by adding common U-Net mistakes, such as common trunks between independent pulmonary veins (PVs) or holes and bumps in the LA surface and body. The final algorithm consisted of two different DAE models, one for bumps and holes removal and a another to separate mistakenly joined PVs, which increased the segmentations' mean dice between ROI and background classes up to 98.4% and 95.4%, respectively. Both numerical and visual results were analyzed. In conclusion, this work proves denoising autoencoders to be an effective post-processing step to refine the output of cardiac anatomical segmentations and improve the accuracy. Since the proposed approach is independent from the U-Net model, the results suggest that the it can lead to more reliable and efficient diagnoses in clinical applications that require accurate cardiac segmentations not only for LA but also for other anatomical regions.