Estimation of the Cardiac Ejection Fraction from image statistics

The Cardiac Ejection Fraction (EF) is an essential criterion in cardiovascular disease prognosis. In clinical routine, EF is often computed from manually or automatically segmenting the Left Ventricle (LV) in End-dyastole and Endsystole frames, which is prohibitively time consuming and needs user interactions. In this paper, we propose a method to minimize user effort and estimate the EF directly from image statistics via machine-learning techniques, without the need for comprehensive segmentations of all the MRI images in a subject dataset. From a user-provided segmentation of a single image, we build a statistic based on the Bhattacharyya coefficient of similarity between image distributions for each of the images in a subject dataset (200 images). We demonstrate that these statistical features are non-linearly related to the LV cavity areas and therefore can be used to estimate the EF. We used Principal Component Analysis (PCA) to reduce the dimensionality of the features and areas. Then, an Artificial Neural Network (ANN) was used to predict the LV cavity areas from the dimension-reduced features. The EF is finally estimated from the obtained areas.