Shape Discrimination of Healthy and Diseased Cardiac Ventricles using Medial Representation

Our work aims at investigating the suitability of the medial representation method to model and analyze shape and shape differences between healthy and diseased hearts. For this experimental study, we use MRI short axis scans of 11 healthy volunteers (age: 50±10) and 5 patients (age: 57±11) with dilativ cardiomyopathy. Controlled semi- automated segmentation provides labels, which are used for the modeling process. To evaluate the model to image accuracy the similarity index (SI), the mean Euclidean distance (ED), and the Hausdorff distance (HD) are calculated. A very high SI (SI > 0.9) for the ventricles is achieved. The mean ED is less than two times the voxel size (1.56 mm) and the HD values for both chambers are in the range of 4.8±3 mm. Applying extended principal component analysis (PCA) on all 16 subjects reveals the distribution of the individual shapes, where the first two PC cover more than 40%, and the first ten PC cover 95% of the shape space. The components show meaningful modes of variation, whereas the healthy and diseased hearts are clustered in the first two components. This preliminary result using the medial based approach promises to discriminate at least globally between healthy and diseased hearts.