Cardiac disease prediction from spatio-temporal motion patterns in cine-MRI

This work presents a motion descriptor that allows to predict cardiac diseases from the quantification of motion patterns in cine-MRI sequences. The proposed approach starts by automatically detecting regions of interest (RoIs) from a fast dense Hough template representation. Then, over the selected RoIs is computed a dense optical flow with the ability to characterize local large displacements. A spatial regional segmentation of RoIs was carried out by using a circular template, where each subregion is characterized by dense orientation flow histograms. The set of subregions motion histograms form the motion descriptor that is mapped to a previously trained support vector machine (SVM) to predict cardiac pathologies. The proposed descriptor was validated on a public dataset with 45 cine-MRI images from 4 pathologies. In average, the proposed descriptor achieved an accuracy of 70.7% in the task of pathologies recognition from a motion descriptor with length of 1140 scalar elements.