Comparative Evaluation of Two Registration-based Segmentation Algorithms: Application to Whole Heart Segmentation in CT

Statistical shape models learn valid variability from example shapes, making large training sets favorable. Methods for automatic training set generation use transforms obtained by registration to propagate atlas landmarks to new samples. Algorithms based on B-spline transforms and mutual information (MI) were successfully employed for the cardiac anatomy in CT and MRI. For single-modality data, however, computationally less complex algorithms such as Thirion’s Demons can be used, allowing for reduced computation times. We implemented two multi-resolution registration-based segmentation pipelines based on Thirion’s Demons, and MI-driven B-spline transforms, respectively, fixed the parameters, and evaluated their performance in whole heart segmentation of contrasted CT angiography images. The segmentation quality was assessed qualitatively using visual inspection and quantitatively using expert ratings. While the Demons-based algorithm required less computation time, the results of the B-splinebased pipeline were in better agreement with the tested data and achieved a higher expert score (3.33± 0.51 compared to 2.19± 0.45). We found registration using B-spline transforms and MI to be favorable, as the application is not time-sensitive. KeywordsRegistration, Statistical Shape Model, Automatic Segmentation, Heart Modelling, CT