Massive Dimensionality Reduction for the Left Ventricular Mesh

Statistical emulation is a promising approach for the translation of cardio-mechanical modelling into the clinical practice. However, a key challenge is to find a low-dimensional representation of the heart, or, for the specific purpose of diagnosing the risk of heart attacks, the left-ventricle of the heart. We consider the problem of dimensionality reduction of the left ventricular mesh, in which we investigate three classes of techniques: principal component analysis (PCA), deep learning (DL) methods based on auto-encoders, and a parametric model from the cardio-mechanical literature. Our finding is that PCA performs as well as the computationally more expensive DL methods, and both outperform the state-of-the-art parametric model.