In this paper we present a new method for fully automatic left ventricle segmentation from 4D cardiac MR datasets. To deal with the diverse dataset, we propose a fully automatic machine learning approach using two layers of spatio-temporal decision forests with almost no assumptions on the data or segmentation problem. We introduce 3D spatio-temporal features to classication with decision forests and propose a method for context aware MR intensity standardization and image alignment. The second layer is then used for the nal image segmentation. We present our first results on the STACOM LV Segmentation Challenge 2011 validation datasets.