Synthesis of Realistic Subcortical Anatomy with Known Surface Deformations

Abstract

Statistical shape analysis is an actively studied field which has been playing an important role in the morphometric analysis of anatomical structures. Many algorithms have been proposed to compare the shape of anatomical structures across populations and a number of theoretical and technical advances have been developed on the subject. Unfortunately, there is very little work on the evaluation and validation of these new techniques. In contrast, comprehensive and objective evaluation frameworks have been designed for closely related fields such as image segmentation, registration or denoising, and play an important role in evaluating the quality of newly proposed algorithms. One possible reason may lie in the difficulty of generating a large set of synthetic shapes which are anatomically realistic, and another set of similar shapes with statistical difference from the previous group. In this work, we try to solve this shape synthesis problem, the outcome of which could be used for the purpose of evaluating shape analysis algorithms. Our method is based on two key components. First, a manifold learning algorithm is applied to a set of known control shapes, giving us the ability to generate an infinite number of new shapes. Second, a multi-scale clustering algorithm allows us to apply known, consistent and realistic deformations to input shapes surfaces. Our framework can thus provide arbitrarily many shapes (of the same organ/structure) with “ground truth” deformations. We present several synthetic, but anatomically derived, shape sets of varying complexity.