IMAGE-BASED DETECTION OF CORPUS CALLOSUM VARIABILITY FOR MORE ACCURATE DISCRIMINATION BETWEEN DYSLEXIC AND NORMAL BRAINS Ahmed Elnakib , Ayman El-Baz , Manuel F. Casanova , Georgy Gimel’farb , and Andrew E. Switala BioImaging Laboratory, Bioengineering Department, University of Louisville, Louisville, KY, USA. Department of Psychiatry and Behavioral Science, University of Louisville, Louisville, KY, USA. Department of Computer Science, University of Auckland, Auckland, New Zealand. ABSTRACT Dyslexia severely impairs learning abilities of children, so that im- proved diagnostic methods are needed. Neuropathological studies have revealed an abnormal anatomy of the Corpus Callosum (CC) in dyslexic brains. We propose a new approach to quantitative analysis of three-dimensional (3D) magnetic resonance images (MRI) of the brain that ensures a more accurate quantification of anatomical differences between the CC of dyslexic and normal subjects. It consists of three main processing steps: (i) segmenting the CC from a given 3D MRI using the learned CC shape and visual appearance; (ii) extracting a centerline of the CC; and (iii) cylindrical mapping of the CC surface for its comparative analysis. Our experiments revealed significant differences (at the 95% confidence level) between 14 normal and 16 dyslexic subjects in four anatomical divisions, i.e. splenium, rostrum, genu and body of their CCs. Index Terms— Segmentation, modeling, corpus callosum, dyslexia. 1. INTRODUCTION Developmental brain disorders belong to one of the most interesting and challenging research areas in modern neuroscience. Dyslexia and autism are two of the most complicated developmental brain disor- ders that affect children’s behavior and learning abilities. Dyslexia leads to the failure to develop age-appropriate reading skills in spite of the normal intelligence level and adequate reading instructions [1], whereas autism is characterized by qualitative abnormalities in behav- ior and higher cognitive functions [2]. Multiple studies during the past decade have identified different brain structures involved in the abnormal neuro-development associated with dyslexia. For example, Casanova et al. [3] and Eliez et al. [4] demonstrated a reduction in the gyral index (i.e. the ratio of the pial surface’s contour to the convex hull of the brain surface) of dyslexic patients. They suggest any gyral abnormality resides in folding rather than thickness of the cortex. This paper develops a new framework for analyzing the surface of CC for normal and dyslexic subjects. The goal is to identify whether or not the CC involved in the abnormal neural development is associated with dyslexia. The CC is the largest fiber bundle connecting the left and the right cerebral hemispheres in the human brain. Since human reading skills are highly affected by the impaired communication between the hemi- spheres, the analysis of the midsagittal of the CC for dyslexic subjects has been proposed in [5–9]. The CC has been traced from a midsagittal MRI slice either manually [5–8] or with a software package [9], and the statistical difference analysis had been applied to find out which part in the CC contributes significantly to identification of dyslexic brains. Plessen et al. [8] computed the midsagittal CC mean shape of both dyslexic and normal brains and noticed that the 2D CC body length can discriminate between the dyslexic and normal subjects. To the best of our knowledge, all the previous works have focused on analyzing a 2D cross section of the midsagittal of the CC although this is insufficient for detecting the whole anatomic variability of the CC of dyslexic subjects. That the known works exploit only the 2D analysis of the CC is the main motivation behind our work. To en- sure a complete 3D analysis, the whole CC surface (traced from all the slices in which the CC appears) is mapped onto a cylinder in such a way as to compare more accurately various autistic and normal CC. Our cylindrical mapping has been inspired by the functional conformal mapping [10]. Similar to the conformal mapping, it is a bijective (one- to-one) transformation and preserves angular relationships between the points. For these reasons, the conformal mapping was recently consid- ered an efficient technique for surface matching [11] and visualization of various anatomic structures [12]. The paper is organized as follows. Section 2 overviews in brief our CC segmentation using a learned soft CC shape model and an identifi- able joint Markov-Gibbs random field (MGRF) model of 3D MRI and “object–background” region maps. Similar approaches have already been successful in segmenting various 2D MRI and CT objects (see e.g. [13, 14]). Our current algorithm has been modified to account for specific properties of the 3D CC. A 3D shape is described in our mod- ification with a probabilistic model rather than conventional distance map. Section 3 details the extraction of the centerline of the segmented CC by solving the Eikonal equation. In contrast to the known 2D so- lutions (see e.g. [13]), the proposed process evolves in the 3D space in order to detect 3D points of the maximal curvature. The cylindrical mapping of the CC after finding its centerline is described in Section 4. Experimental results and conclusions are presented in Section 5. 2. SEGMENTATION OF CORPUS CALLOSUM USING A SHAPE MODEL AND A JOINT MGRF MODEL OF 3D MRI Let , , and be a set of integer gray levels, a set of object (“ob”) and background (“bg”) labels, and a unit interval, respectively. Let a 3D arithmetic grid support grayscale MRI , their binary region maps , and probabilistic shape model . The shape model allows for registering (aligning) 3D brain MRI. The co-registered 3D MRI and their region maps are modeled with a joint MGRF specified by a probability distribution where is an uncondi- tional Gibbs distribution of co-registered region maps, is a conditional distribution of the MRI signals given the map, and is a conditional distribution of the prior shape of the CC given the map. We focus on accurate identification of spatial voxel interactions in , voxel-wise distributions of intensities in , and prior distribution of the shape of the CC in for co-aligned 3D MR images. The probabilistic 3D shape model is learned from a train- ing set of manually segmented and co-aligned images. To perform the initial CC segmentation, every given MRI is aligned to one of the training images. The shape model provides the voxel-wise object and background probabilities being used, together with the conditional im- 109 978-1-4244-4126-6/10/$25.00 ©2010 IEEE ISBI 2010