Graph cut segmentation of the right ventricle in cardiac MRI using multi-scale feature learning

It is important to segment the right ventricular structure on MR images for the quantitative evaluation of cardiac function. Some factors such as the distinct cavity border and the variable crescent shape bring the challenge to this task. To overcome these difficulties, a segmentation framework integrating graph cut with multi-scale feature learning is presented. Firstly, multi-scale image patches extracted from the training set are used to perform the supervised feature learning by stacked sparse auto-encoder. Secondly, the right ventricular structure is roughly located utilizing the estimated probability map for the testing set. Finally, a regularized energy function based on graph cut is optimized for segmentation refinement. A number of comparative experiments demonstrate the superiority of the proposed method in terms of segmentation accuracy.