Date on Master's Thesis/Doctoral Dissertation


Document Type

Doctoral Dissertation

Degree Name

Ph. D.


Electrical and Computer Engineering

Committee Chair

El-Baz, Ayman Sabry

Author's Keywords

Image modeling; MRI; Segmentation; Left ventricle; Dyslexia; Computer aided diagnosis


Diagnostic Imaging; Dyslexia--Diagnosis


Detecting abnormalities in two-dimensional (2D) and three-dimensional (3D) medical structures is among the most interesting and challenging research areas in the medical imaging field. Obtaining the desired accurate automated quantification of abnormalities in medical structures is still very challenging. This is due to a large and constantly growing number of different objects of interest and associated abnormalities, large variations of their appearances and shapes in images, different medical imaging modalities, and associated changes of signal homogeneity and noise for each object. The main objective of this dissertation is to address these problems and to provide proper mathematical models and techniques that are capable of analyzing low and high resolution medical data and providing an accurate, automated analysis of the abnormalities in medical structures in terms of their area/volume, shape, and associated abnormal functionality. This dissertation presents different preliminary mathematical models and techniques that are applied in three case studies: (i) detecting abnormal tissue in the left ventricle (LV) wall of the heart from delayed contrast-enhanced cardiac magnetic resonance images (MRI), (ii) detecting local cardiac diseases based on estimating the functional strain metric from cardiac cine MRI, and (iii) identifying the abnormalities in the corpus callosum (CC) brain structure—the largest fiber bundle that connects the two hemispheres in the brain—for subjects that suffer from developmental brain disorders. For detecting the abnormal tissue in the heart, a graph-cut mathematical optimization model with a cost function that accounts for the object’s visual appearance and shape is used to segment the the inner cavity. The model is further integrated with a geometric model (i.e., a fast marching level set model) to segment the outer border of the myocardial wall (the LV). Then the abnormal tissue in the myocardium wall (also called dead tissue, pathological tissue, or infarct area) is identified based on a joint Markov-Gibbs random field (MGRF) model of the image and its region (segmentation) map that accounts for the pixel intensities and the spatial interactions between the pixels. Experiments with real in-vivo data and comparative results with ground truth (identified by a radiologist) and other approaches showed that the proposed framework can accurately detect the pathological tissue and can provide useful metrics for radiologists and clinicians. To estimate the strain from cardiac cine MRI, a novel method based on tracking the LV wall geometry is proposed. To achieve this goal, a partial differential equation (PDE) method is applied to track the LV wall points by solving the Laplace equation between the LV contours of each two successive image frames over the cardiac cycle. The main advantage of the proposed tracking method over traditional texture-based methods is its ability to track the movement and rotation of the LV wall based on tracking the geometric features of the inner, mid-, and outer walls of the LV. This overcomes noise sources that come from scanner and heart motion. To identify the abnormalities in the CC from brain MRI, the CCs are aligned using a rigid registration model and are segmented using a shape-appearance model. Then, they are mapped to a simple unified space for analysis. This work introduces a novel cylindrical mapping model, which is conformal (i.e., one to one transformation and bijective), that enables accurate 3D shape analysis of the CC in the cylindrical domain. The framework can detect abnormalities in all divisions of the CC (i.e., splenium, rostrum, genu and body). In addition, it offers a whole 3D analysis of the CC abnormalities instead of only area-based analysis as done by previous groups. The initial classification results based on the centerline length and CC thickness suggest that the proposed CC shape analysis is a promising supplement to the current techniques for diagnosing dyslexia. The proposed techniques in this dissertation have been successfully tested on complex synthetic and MR images and can be used to advantage in many of today’s clinical applications of computer-assisted medical diagnostics and intervention.