In this paper we present a method for fully automatic left atrium segmentation from 3D cardiac magnetic resonance datasets. We propose a machine learning approach using decision forests that requires very few assumptions on the segmentation problem. First, we extract the blood pool using a simple thresholding technique. Then, we learn to separate the left atrium from other structures in the image by using context-rich features applied on images enhanced with a multi-scale vesselness filter and transformed to measure distance to blood pool surface. We present our results on the STACOM LA Segmentation Challenge 2013 validation datasets.