Integrating dimension reduction with mean-shift clustering for biological shape classification

Quantitative shape analysis is required in a broad range of biological studies. Mean-shift clustering provides a powerful approach for automated biological shape classification because it is a nonparametric clustering technique that does not impose artificial constraints on the number and distributions of the shape classes. However, the high-dimensionality of the shape space often causes significant performance deterioration in kernel density estimation in mean-shift clustering. To address this problem, we developed a dimension reduction approach that preserves the geometrical structure of the shape space while allowing a significant acceleration of mean-shift clustering computation by more than one order of magnitude. We validated performance of the algorithm on a generic shape dataset and then used the algorithm to analyze morphology of axonal mitochondria in neurons.