(A) template model based shape morphometry

Abstract

Shape morphometry based on the 3D shape model of human organs has emerged as a preferred tool to investigate their morphological changes with respect to pathological processes (e.g., neurodegenerative diseases and cancers). Various methods have been proposed for the shape morphometry, but it is still challenging to estimate accurate surface boundaries against rough and noisy segmentations and to achieve good anatomical correspondence between the individual models. In addition, the sensitivity and comparability of the geometric measures, which quantify the shape differences between subjects, need to be validated and guaranteed for the comparison of the individuals` shapes with anatomical knowledge. In order to address these issues in the shape morphometry, we introduce a template model-based approach on the shape modeling and its quantification. In our approach, the template model is defined as a pre-defined layout of their shape representations and anatomical landmarks for the shape feature measurement. The template model includes not only the surface mesh to encode the generic shape information of the target structure but also anatomical landmarks to determine the spatial normalization of the shape models and the baseline for measuring the geometric differences between corresponding primitives in the individualized models. For the template model-based morphometry, we develop computational methods for the template model construction, the template model adaptation to individual subjects, and the template model-based shape quantification with a focus on the morphology analysis of the hippocampus and the third ventricle. More specifically, we first propose a template model representation based on the multi-level neighborhood and the Laplacian coordinates and a progressive surface model deformation minimizing the geometric distortion while reconstructing the target shape via a non-rigid model deformation. In the study on the third ventricle, we extend the definition of the template model to include the midsagittal of the brain as an anatomical basis for the shape morphometry to investigate the morphological changes of the third ventricle in relation to brain atrophy in the left and right hemispheres separately. In this dissertation, we present the experiments showing (1) accuracy and robustness of the proposed methods in smooth surface reconstruction with the distortion minimization while filtering out high-frequency noises in input volumes, (2) its sensitivity in detecting significant shape differences with respect to the neurodegenerative diseases (mild cognitive impairment and Alzheimer`s disease) and cognitive ability, (3) its robustness in constructing the anatomical shape correspondence between the individualized models, and (4) its reproducibility and consistency in quantifying the shape characteristics of target structures using the shape models and feature descriptors. In the experiments, we compared the performance of the proposed method with other well-known shape modeling methods: SPHARM-PDM, ShapeWorks and LDDMM volume registration with template injection. In addition, we employed various metrics of shape similarity, surface roughness, volume and shape deformity of the structures to quantitatively investigate the performance (e.g., accuracy, robustness, sensitivity and denoising effect) of the proposed method. The experimental results showed that the proposed method performed better than others in generating smoothen surfaces with less volume differences and better shape similarity to input volumes. The statistical analyses with clinical variables also showed that it was sensitive enough to detect subtle shape differences of the brain structures.