Research

Computational Analysis of Brain Images: Towards a Useful Tool in Clinical Practice

In DTU Compute PHD-2015, 2016

Read the article

Abstract

Due to its excellent soft tissue contrast and versatility, magnetic resonance imaging (MRI) has become arguably the most important tool for studying the structure and disorders of the human brain. Although in recent years tremendous advances have been made in automatic segmentation of brain MRI scans, many of the developed methods are not readily extendible to clinical applications due to the variability of clinical MRI data and the presence of pathologies, such as tumors or lesions. Thus, clinicians are forced to manually analyze the MRI data, which is a time consuming task and introduces rater-dependent variability that reduces the accuracy and sensitivity of the results. The goal of this PhD-project was to enlarge the scope of the automatic tools into clinical applications. In order to tackle the variability of the data and presence of pathologies, we base our methods on Bayesian generative modeling, which combines detailed prior models of the human neuroanatomy and pathologies with models of the MRI imaging process. This approach allows us to describe the observed MRI data in a principled manner, and to integrate explicit models of different disease effects and imaging artifacts into the framework when needed. This thesis presents an introduction to the theory behind the generative modeling approach, and an overview of the main results. The first part concentrates on segmenting different neuroanatomical structures in MRI scans of healthy subjects, and the second part describes how this framework can be extended with models of brain lesions. This results in a set of fast, robust and fully automatic tools for segmenting MRI brain scans of both healthy subjects and subjects suffering from brain disorders such as multiple sclerosis. Having access to quantitative measures of both lesions and the surrounding structures opens up avenues for clinicians to study the effect of these type of disorders on the full brain anatomy. This could potentially help in discovering sensitive biomarkers for early diagnosis and tracking of disease development.