A Unified Statistical/Deterministic Deformable Model for LV Segmentation in Cardiac MRI

We propose a novel deformable model with statistical and deterministic components for LV segmentation in cardiac magnetic resonance (MR) cine images. The statistical deformable component learns a global reference model of the LV using Principal Component Analysis (PCA) while the deterministic deformable component consists of a finite-element deformable surface superimposed on the reference model. The statistical model accounts for most of the global variations in shape found in the training set while the deterministic skin accounts for the local deformations consistent with the detailed image features. Intensity gradient-based image forces are applied to the model to segment and reconstruct LV shape. We validate our model on the MICCAI Grand Challenge dataset using leave-one-out training. Comparing the automated segmentation to the manual segmentation yields a Mean Perpendicular Distance (MPD) of 3.65 mm and a Dice coefficient of 0.86.