Medical image segmentation is the process of segmenting/ sectioning out a particular structure of interest from an entire image, which is obtained from an imaging modality such as MRI or CT. The segmentation procedure used often depends on different factors such as the imaging modality, the properties of the structure of interest and the computational performance required. The problem of image segmentation is a widely explored topic in the domain of medical image processing. It makes the study of complex structures easier, which in turn helps immensely in better diagnosis and treatment planning. In this work, the aim is to study the performance of five different approaches for segmenting five different structures of the human brain in a T1 MR image. These methods make use of information from already segmented reference images to perform segmentation on the input and hence are classified as multi-atlas (multiple references) based techniques. They treat the entire brain volume as a group of patches (made of individual voxels) and perform segmentation by operating at the patch level and hence are called the patch based methods. The Dice coefficient is used as a measure to evaluate segmentation performance by each of these methods. Through this analysis, the objective is to implement, understand and analyze each of these methods and also identify their shortcomings.

The finite element method (FEM) is a way of solving partial differential equation on a spatially (and temporarily) discretized model. It is widely used in medical applications. In this research, we develop a tool to calculate the response to an impact on human head and brain. We use the finite element method to model the solid tissue as linear elastic materials and the fluid-filled (1) as impressible viscous fluid. The geometric (2) model is derived from Magnetic Resonance Imaging (MRI) data of a real subject. To solve the nonlinear dynamic problem, we use the Newton-Raphson method, and employ parallelization to speed up the computation. Example simulations demonstrate the validation of the model.

The linear model (LM) is typically used to analyze the relationship between imaging data and demographic/cognitive parameters. Each imaging measurement is considered as independent in this model. Recent neuroimaging studies collect time-series data, for which the assumption of independence is invalid. Instead, we use the linear mixed model (LMM) that gives us population effects and subject effects as regression coefficients. A downside of LMM is the computation burden. The purposes of this research are: (1) to develop the tools for analyzing large data, (2) to interpret results, and (3) to demonstrate how to use subject wise information.

We focused on a large dataset of diffusion weighted MR images. 730 images and demographic/cognitive tests were acquired in 176 subjects by the Alzheimer's Disease Neuroimaging Initiative (ADNI). Our code can successfully analyze the 30GB dataset within one day. We found that the population effects coefficients of LMM are roughly similar to the coefficients estimated by LM, and the statistical significance of regressors in the LMM is typically lower than in the LM. We demonstrated that subject wise information can be used to determine an onset of deterioration for each healthy control. Results suggest that this parameter is helpful to model the general aging process.

Traumatic brain injury is the main cause of disability and death in children and infants. This thesis focuses on establishing the consequences of an external force or impact on brain tissues. FE model has been introduced to investigate the consequences of impact force, either dynamic or static, onto the infant brain. Deformable meshes are involved in FE models to generate time-dependent pressure change in various brain tissues. Results demonstrate that our finite element model is capable of investigating brain tissue pressure transmission within certain impact force and can be used for simulation of minor brain injury in infants.

Biomaterials, including brain tissue, usually have nonlinear material properties. In this project, a human head simulator using the finite element method (FEM) was improved by adding an implementation of a nonlinear material model into it. The Ogden model is chosen to be implemented in this project because of its good accuracy to model the brain tissue. This work aims at improving the validity of impact simulation tests. Results show that the nonlinear model was successfully implemented into the simulator and a difference was discovered when conducting impact simulation. In the simulation using a nonlinear model, the deformation of brain tissue is larger than the deformation using a linear model.

Spherical Mean technique (SMT) is a novel method of quantifying the diffusion properties of the nerve fibers bundles in the central nervous system. It does this by calculating the spherical mean of the diffusion signal and fitting it to a parametric equation to obtain per voxel diffusion coefficients. We used Expectation - Maximization to obtain Gaussian Mixture Models (GMM) to find distinct clusters in per voxel coefficient space. We found that the diffusion properties of all the white matter fibers were clustered into a single Gaussian distribution in 867 brain volume samples. This implies that the diffusion properties of the white matter fibers are relatively homogeneous. Then, we checked this result by comparing the clusters obtained using GMM with tissue classification outputs obtained by clustering Fractional Anisotropy (obtained using Diffusion Tensor modeling), T1 weighted image intensity and B0 image intensity for 867 brain volume samples; we observed that the specific clusters of per voxel diffusion coefficients obtained using GMM represent specific tissue types (grey matter fibers, white matter fibers, cerebrospinal fluid). Since the parameters derived from SMT represent the physical diffusion properties that are independent of microscopic fiber orientation and the distribution of diffusion coefficients of white matter can be modeled by a single Gaussian distribution, we can conclude that the diffusion properties of all white matter fiber are homogeneous.

Storage of adipose tissue is a trait belonging to all humans, historically borne from the necessity to survive extended periods of famine. Yet, there is clear inter-individual variability in our propensity to accumulate excess fat. In addition, this variability is present as early as infancy and persists into adulthood. In adults, the regulation of energy homeostasis (balance) by the brain is evident not only through signaling provided by peripheral systems to the hypothalamus, such as fat (leptin) and the stomach (ghrelin), but also by the downstream cognition required to decide what, when and how much to eat. Magnetic Resonance Imaging has been instrumental in identifying the brain regions and circuitry involved in food-related behaviors as well as the identification of brain differences between obese and normal weight adults. It is, however, unclear whether the observed differences in brain regions and circuitry in obese relative to normal weight individuals are a cause, consequence, or both, of the obese state. Moreover, relatively little is known about the developmental ontogeny of these food-related brain regions and circuitry, particularly during the period of intrauterine development (when the postnatal obesogenic environment could not yet have affected this circuitry), and its prospective role in shaping propensity for childhood obesity. Knowledge of early life brain structure and function in the context of energy imbalance would contribute to an improved understanding of propensity for obesity, early identification of at-risk individuals, and intervention targets for primary prevention. This thesis has addressed this fundamental knowledge gap by using an imaging based approach to: 1) quantify and describe early life patterns of fat deposition, 2) shed new insights on perinatal brain development, and 3) identify prospective associations between interoceptive, reward, and gustatory properties of the newborn brain and subsequent early life fat gain.

From ancient times, there have been conflicting views of the significance of the brain. Such different opinions over the years show how little was known of the brain’s anatomy. The idea that our brains have a common basic structure, although it may seem rather straightforward, was not developed until 200 years ago when scientists first began to give names to structures. If all brains have a common structure, a data-driven study of different brain surfaces should have also similar results. In this thesis, a dataset of 100 brains is studied to see if a consistent partitioning among them can be found. Its aim it to asses whether the lobe partitioning (frontal, parietal, temporal and occipital) is also supported by data and, if not, to find a good data-driven partitioning of the brain. To do so, surface brain meshes are generated from MRI data and then treated with a partitioning software to segment them into different parts. The consistency among brains is analyzed with both registration and anatomical automatic labeling (AAL). It can be extracted from our analysis that the best data-based partitioning of our brains corresponds to the 4-partitioned meshes for both methods of consistency used, registration and anatomical labeling.