Automatic and robust segmentation of the left ventricle (LV) in magnetic resonance images (MRI) has remained challenging for many decades. With the great success of deep learning in object detection and classification, the research focus of LV segmentation has changed to convolutional neural network (CNN) in recent years. However, LV segmentation is a pixel-level classification problem and its categories are intractable compared to object detection and classification. Although lots of CNN based methods have been proposed for LV segmentation, no robust and reproducible results are achieved yet. In this paper, we try to reproduce the CNN based LV segmentation methods with their disclosed codes and trained CNN models. Not surprisingly, the reproduced results are significantly worse than their claimed accuracies. We also proposed a fully automated LV segmentation method based on slope difference distribution (SDD) threshold selection to compare with the reproduced CNN methods. The proposed method achieved 95.44% DICE score on the test set of automated cardiac diagnosis challenge (ACDC) while the two compared CNN methods achieved 90.28% and 87.13% DICE scores. Our achieved accuracy is also higher than the best accuracy reported in the published literatures. The MATLAB codes of our proposed method are freely available on line.