Pattern analysis of morphometric features from biomedical image data.

摘要

Pattern analysis is very useful in medical imaging for analyzing biomedical images and detecting important imaging biomarkers. However, the direct application of traditional pattern analysis techniques may not achieve desirable performance for some biomedical images, because those techniques usually do not consider the characteristics of the morphometric features from these biomedical image data. This dissertation aims to design new techniques for extracting and analyzing morphometric features from biomedical image data and validate the effectiveness of the proposed methods in practical medical studies. In the first part, this dissertation presents two general frameworks for extracting morphometric features from 3D closed surfaces and from 3D disk-like surfaces. For 3D closed surfaces with spherical topology, spherical harmonic (SPHARM) shape modeling framework is implemented to represent this type of shape. To deal with a group of 3D surfaces with disk-like topology, a new computational framework integrating a set of effective surface registration methods is proposed to form a unified surface based morphometry processing pipeline. Both frameworks combining general linear regression and random field theory are applied to real medical studies. The effectiveness of the two frameworks is demonstrated by identified regional shape changes related to certain conditions. In the second part, the dissertation focuses on more advanced techniques for morphometric feature analysis. First of all, two multivariate sparse models, Elastic net (EN) and sparse canonical correlation analysis (SCCA), are employed to examine the genetic effects in hippocampal shape changes in an Alzheimer's disease (AD) study. The two models show great power to reveal complex relationship between single nucleotide polymorphisms (SNPs) and hippocampal shape features. Secondly, an efficient sparse Bayesian multi-task learning algorithm is proposed to adaptively learn and exploit the dependence relation among multiple responses by modeling the inter-vector correlations in the regression coefficient matrix. The application of this algorithm to predicting cognitive performance from MRI measures in an AD study demonstrates the proposed algorithm has superior prediction performance over multiple state-of-the-art competing methods. Finally, a more advanced sparse Bayesian learning algorithm jointly exploiting both the inter-vector and intra-block correlations in the regression coefficient matrix is designed and applied to the same AD study of predicting cognitive scores. Different from existing sparse algorithms, the new algorithm has the ability to model the response as a nonlinear function of the predictors by extending the predictors matrix with block structures. Experimental results show that this algorithm not only achieves better prediction performance than its predecessor and other competing algorithms, but also effectively identifies biologically meaningful patterns. In the last part, the dissertation provides a comparative evaluation of typical analysis methods and morphometric features. Six typical/classical regression algorithms are compared in the task of predicting cognitive performances from four types of hippocampal imaging measures. The comparison results demonstrate the multivariate sparse Bayesian learning exploiting the correlation structures is a valuable framework in discovering biomarkers related to cognitive performance and subfield imaging measures yield the most powerful and stable prediction rates across all the algorithms.