Deep Convolutional Neural Networks for left ventricle segmentation.

Abstract

Left ventricle (LV) segmentation is crucial for quantitative cardiac function analysis. Manual segmentation of the endocardium and epicardium is highly cumbersome; physicians limit delineation to the end-diastolic and end-systolic phases. A fully automated system could provide an analysis of cardiac morphology for all phases in a much shorter time. Most of the current LV segmentation methods are semi-automated and require error prone manual initialization. A fully-automated LV segmentation method would expedite the functional analysis of the LV, reduce subjectivity and improve patient experience. We automatically segment the LV wall in cardiac MRI images with a Deep Convolutional Neural Network (DCNN). This algorithm first calculates the probability of a pixel belonging to the LV wall or background and then generates a label based on those probabilities without manual initialization. We then compare these results to the results obtained with another DCNN initialization method using Gabor filters. With Gabor DCNN we obtain an accuracy of 0.97, specificity of 0.984, sensitivity of 0.841 and mean accuracy of 0.902. This shows that Gabor filters perform better than random filters in the DCNN for LV segmentation.