Direct Learning Left Ventricular Meshes from CMR Images

Biomechanical studies of the left ventricle (LV) typically rely on a mesh of finite element nodes for a discrete representation of the LV geometry, which is used in an approximate numerical solution of the cardio-mechanical equations based on finite-element discretisation. This is typically done by first manually annotating cardiovascular magnetic resonance (CMR) scans, second creating a preliminary mesh, third manually correcting the mesh to account for motion. The whole process requires specialist knowledge, is time consuming and prone to human error, which prohibits its common adoption in the clinics. We propose to overcome these shortcomings by applying statistical pattern recognition techniques to CMR images. In particular, we train a convolutional neural network (CNN) to predict the LVM via learning its principal component representation directly from CMR scans. As a useful side-product we obtain a low-dimensional representation of the LVM, which is of interest for surrogate models (emulators) of the myocardium constitutive models.