Online Relational Manifold Learning for Multiview Segmentation in Echocardiography

Accurate delineation of the left ventricle (LV) endocardial border in echocardiography is of vital importance for the diagnosis and treatment of heart disease. Effective segmentation of the LV is challenging due to low contrast, signal dropout and acoustic noise. In the situation where low level and region-based image cues are unable to define the LV boundary, shape prior models are critical to infer shape. These models perform well when there is low variability in the underlying shape subspace and the shape instance produced by appearance cues does not contain gross errors, however in the absence of these conditions results are often much poorer. In this paper, we first propose a shape model to overcome the problem of modelling complex shape subspaces. Our method connects the implicit relationship between image features and shape by extending graph regularized sparse nonnegative matrix factorization (NMF) to jointly learn the structure and connection between two low dimensional manifolds comprising image features and shapes, respectively. We extend conventional NMF learning to an online learning-based approach where the input image is used to leverage the learning and connection of each manifold to the most relevant subspace regions. This ensures robust shape inference and a shape model constructed from contextually relevant shapes. A fully automatic segmentation approach using a probabilistic framework is then proposed to detect the LV endocardial border. Our method is applied to a diverse dataset that contains multiple views of the LV. Results show the effectiveness of our approach compared to state-of-the-art methods.