Detecting and refurbishing ground truth errors during training of deep learning-based echocardiography segmentation models