Whole Heart Segmentation of Cardiac MRI Using Multiple Path Propagation Strategy

Automatic segmentation of cardiac MRI is an important but challenging task in clinical study of cardiac morphology. Recently, fusing segmentations from multiple classifiers has been shown to achieve more accurate results than a single classifier. In this work, we propose a new strategy, MUltiple Path Propagation and Segmentation (MUPPS), in contrast with the currently widely used multi-atlas propagation and segmentation (MAPS) scheme. We showed that MUPPS outperformed the standard MAPS in the experiment using twenty-one in vivo cardiac MR images. Furthermore, we studied and compared different path selection strategies for the MUPPS, to pursue an efficient implementation of the segmentation framework. We showed that the path ranking scheme using the image similarity after an affine registration converged faster and only needed eleven classifiers from the atlas repository. The fusion of eleven propagation results using the proposed path ranking scheme achieved a mean Dice score of 0.911 in the whole heart segmentation and the highest gain of accuracy was obtained from myocardium segmentation.