Accurate localization of the left ventricle (LV) boundary from echocardiogram images is of vital importance for the diagnosis and treatment of heart disease. Statistical shape models such as active shape models (ASM) have been commonly used to perform automatic detection of this boundary. Such models perform well when there is low variability in the underlying shape subspace and an accurate initialization can be provided, however in the absence of these conditions results are often much poorer. In the case of LV echocardiogram images, such variability is often encountered in patients with abnormal LV function. In this paper we propose a fully automatic segmentation technique using deep learning in a Bayesian nonparametric framework. Our model uses a dynamic statistical shape model comprised of training shapes from select weighted subsets of the feature subspace. Subsets are chosen during the iterative segmentation process according to a latent temporal component allocation variable, determined from joint deep features and LV landmark information using a Dirichlet process mixture model with Chinese restaurant process prior. Testing is performed on a data set comprising images of the LV acquired from patients exhibiting both normal and abnormal LV function, and the results using our technique compared to both the ASM and other state of the art techniques. Results from this testing show an improvement in the LV localization accuracy, particularly when LV function is abnormal.