Groupwise Morphometric Analysis Based on High Dimensional Clustering.

In this paper, we propose an efficient groupwise morphometric analysis to characterize morphological variations between healthy and pathological states. The proposed framework extends the work of Baloch [4] in which a manifold for each anatomy was constructed by collecting lossless [transformation, residual] descriptors with various transformation parameters, and the optimal set of transformation parameters was estimated individually by minimizing group variance. However, full parameter exploration is not desired as it can result in transformation leading to inaccurate anatomical models. In addition, a single fixed template introduces a priori bias to subsequent statistical analysis. In order to overcome these limitations, we use an affinity propagation clustering method to find the spatially close cluster center for each subject. Then, a subject is normalized to the template via the cluster center to restrict our analysis only to those descriptors that reflect reasonable warps. In addition, a mean template is selected by finding a cluster center that minimizes the sum of pairwise shape distance to reduce the fixed template bias. Our method is applied to 2D synthetic data and 3D real Cardiac MR Images. Experimental results show improvement in quantifying and localizing shape changes.