and other different parameters such as the gyrification index, gyral window using the deformable models will be considered. Second, diffusion tensor images are ...
EFFECT OF MINICOLUMNAR DISTURBANCE ON DYSLEXIC BRAINS: AN MRI STUDY Noha El-Zehiry*, Manuel Casanova**, Hossam Hassan* and Aly Farag* * Computer Vision and Image Processing Lab, University of Louisville ** Department of Psychiatry and Behavioral Sciences, University of Louisville
ABSTRACT The minicolumn is generally considered the basic unit of the neocortex in all the mammalian brains. Enlargement of the cortical surface is believed to occur through the addition of minicolumns rather than a single neuron.This study aims at testing the hypothesis that brain developmental disorders can be diagnosed and analyzed in terms of the minicolumnar disturbance. To do this, we propose to correlate the pathological findings in terms of the minicolumnar structure to the MRI findings in terms of volumetric analysis.
used for brain extraction, segmentation and white matter parcellation. Section 4 will show the experimental results. 2. PROBLEM STATEMENT, MOTIVATION AND DATA SET DESCRIPTION
Autism and dyslexia are two of the most complex developmental disorders that affect children. Dyslexia is characterized by the failure to develop age appropriate reading skills despite normal intelligence level and adequate reading instructions [3]. However, autism is a more complex disease characterized by social interaction, language, behavior and cognitive learning.[4] 1. INTRODUCTION One of the major causes of developmental disorders is that some Brain function research suggest that the minicolumn, not indi- parts of the communications network of the brain fails to pervidual neuronal cells, are the basic unit of operation in the brain. form its tasks properly. Hence, most of the previous MRI studFor example, Mountcastle[1] reported ”The effective unit of op- ies [4, 5, 6] focused on investigating morphometric brain changes eration in such a distributed system is not the single neuron and through some certain structures such as the corpus callosum, its axon, but groups of cells with similar functional properties brain stem and other structures. and anatomical connections” . The basic contribution is that it analyzes the morphometric changes The most highly respected textbooks of neurobiology consider in the normal and dyslexic brains through investigating the minithe vertical organization of the cortex a basic concept of neuro- columnar organization rather than structures. The analysis is biology, commonly referring to it as the columnar hypothesis.[2] based on the fact that the failure to develop proper communicaMinicolumns are defined as vertical organization of neurons tion between the different parts of the brain can be caused by arranged in the neocortex. Minicolumns, rather than the neu- the disturbance in the minicolumns. Meanwhile, postmortem rons, are believed to be the basic unit of the brain, i.e., the en- studies have shown that a common feature to both disorders largement of the neocortex, in the phase of the development is a disturbance with the minicolumns. These studies indicate of human brain, occurs through the addition of minicolumns that within the general population, dyslexia and autism exist as rather than single neurons. And,on the contrary, in the aging opposite tail ends within the normal distribution of the miniphase, the neocortex loses whole minicolumns not some ran- columnar widths [7]. This can be illustrated in Fig. 1 domly scattered neurons. Thus, any disturbance in the number, structure or organization of minicolumns will affect the development of the brain and, as a result, its functionality. According to this, we hypothesize that the brain developmental disorders and aging disorders can be diagnosed and analyzed in terms of the analysis of the number as well as the organization of the minicolumns in the neocortex. This project aims to find a link between the pathological studies Autism Dyslexia or the microscopic domain work and a macroscopic framework represented by a non invasive imaging technique. Therefore, the major objective of this paper is to correlate the pathological findings vis-a-vis the minicolumns to the MRI findings represented by the volumetric analysis of the neocortex. Fig. 1. Illustrating the idea of obtaining the moving the outer contour The paper will be organized as follows; Section 2 will explain using the signed distance function the problem, the data set, MRI protocol and the hypothesis that should be proven. Section 3 will explain the methodologies Thus, in dyslexia, the widths of the minicolumns are generally 20
70 Micron
larger than the minicolumns in normal control cases and the brain is proved, pathologically, to have less number of minicolumns. The converse is true in autism. In the normal brain, the minicolumnar interconnectivity is on the order of 1000 [8]. Therefore, the number of connections (white matter) would be highly affected by any change in the number of minicolumns. Therefore, to conclude the previous discussion, the hypothesis is: in dyslexia, the widths of the minicolumns are larger than they are in normals, and hence; the number of minicolumns in the dyslexic patients is less than its corresponding in normal control cases. Less number of minicolumns means much less connections which implies smaller white matter volume, and the contrary occurs in autism. Therefore,the major objective of this study is two fold; first, is to try to prove this hypothesis by proving that the volume of the white matter in dyslexic cases is smaller than its corresponding in normal control cases. Second, to emphasize that the change in the volume of the white matter is resulting from the communication 1 between the minicolumns, thus, the white matter will be parcellated into inner and outer compartments to prove that the increment or decrement in white matter volume should occur in the outer compartment. 2.1. Data Set and MRI Protocol The data set consists of sixteen right handed dyslexic men aged 18 to 40 years, and a group of 14 controls matched for gender, age, educational level, socioeconomic background, handedness and general intelligence. All the subjects are physically healthy and free of history of neurological diseases, head injury. Briefly, all the subjects have exactly the same psychiatric conditions. All images were acquired with the same 1.5 T MRI scanner (GE, Milwaukee, Wisconsin) using a 3-D spoiled gradient recall acquisition in the steady state (time to echo, 5 milliseconds; time to repeat, 24 milliseconds; flip angle, 45; repetition,1; field of view, 24 cm2 ). Contiguous axial slices, 1.5 mm thickness (124 per brain) were obtained and the voxel resolution was 0.9375*0.9375*1.5mm. 3. METHODS The proposed approach to accomplish the the objective consists mainly of four steps; first, brain extraction algorithm is applied to remove all the non brain tissues from the images. Second, brain segmentation is performed to isolate the white matter. Third, white matter parcellation is performed to parcel the white matter into inner and outer compartments and finally, volumetric measures for the whole white matter, as well as, the outer compartment are recorded, and then tests of hypothesis are performed to investigate if there exists a significant difference between the groups or not. The following subsections will provide a brief description to each of the previously mentioned steps. 1 Communication between minicolumns is represented by the U fibers that connects the different parts of the neocortex, since the minicolumns exist in the gray matter and connects through the white matter, then these U fibers exist in the white matter layer that is very close to the gray matter ”Outer Compartment of the White matter”
3.1. Brain extraction and segmentation Brain extraction is being applied to remove all the non brain tissues such as skull and fat. The MRICro software has been utilized to perform skull stripping and extract the brain . Having extracted the brain, Level sets segmentation approach has been utilized for the segmentation. Segmentation partitions the image into regions [9] each belonging to a certain class. In our approach a separate level set function is defined for each class and automatic seed initialization is used. The probability density functions of classes are embedded into the velocity term of each level set equation. The parameters of each one of these density functions are re-estimated at each iteration. The competition between level sets based on the probability density functions stops the evolution of each level set at the boundary of its class region. Let Ω ∈ Rn be open and bounded n-dimensional volume. Let I : Ω → R be the observed p-dimensional image data. We assume that the number of classes K is known. Let pi (I) be the intensity probability density function of class i (assumed to be Gaussian). In this work we associate the mean µi , variance σi2 , and prior probability πi with each class i. In accord to the estimation methods in [9, 10], the model parameters are updated at each iteration. The classification decision is based on Bayes’ decision [10] at point x as follows: i∗ (x) = arg max (πi pi (I(x))). i=1,..,K
(1)
The term (ν) (the moving front directional term) for each point x is replaced by the function νi (x) ([9]) so the velocity function is defined as: Fi (x) = νi (x) − ²κ(x), ∀ i = 1..K. (2) ∗ where νi (x) = −1 if i = i (x) and it is 1 elsewhere. If the pixel x belongs to the front of the class i = i∗ (x) associated to the level set function, the front will expand, otherwise it will contract. Now, we put the level set equation in the general form using the derivative of the Heaviside step function (δα (z)) as follows : ∂φi (x, t) = δα (φi (x, t))(²κ(x) − νi (x))|∇φi (x)|. (3) ∂t The function δα (z) selects the narrow band points around the front κ is the curvature. The proposed approach has been used to classify between the white matter, gray matter and cerebrospinal fluid. The approach has proved to be promising in extracting the white matter. 3.2. White matter parcellation White matter parcellation aims at dividing the white matter into inner compartment and outer compartment: the outer compartment is a strip of the white matter where the connectivity between the minicolumns occurs. This area is extracted by detecting the outer contour of the white matter and moving this contour inwards 2 to obtain a boundary between the inner and outer compartments. The way we see the outer compartment 2 By an arbitrary distance that is sufficiently large to include all the U fibers that connects between the minicolumns
of the white matter or the region of interest that we are looking for is the region of the white matter where the minicolumns communicate to each other, i.e. the region where the axon ends and the synapses communicate to different dendrites to transfer information. Therefore, we have two choices to determine such a region; either to ask a neuroscientist to draw this boundary (between the inner and outer compartments) manually which is very effort and time consuming. Or, to choose a distance that will be sufficiently large so that it is guaranteed that the communications between the minicolumns occur through it. Bearing in mind that if the chosen distance is larger than the actual one ”that could be determined by a neuroscientist”, this will just add a bias to the volumes but the differential comparison will still be valid. For this purpose, level set methods are used to detect the outer contour of the white matter, and then the signed distance map φ(x) of the contour Γ defined as [11] − → x ) if − x is inside Γ d(→ → − 0 if x x ∈ Γ φ(x) = (4) → − -d(− x ) if → x is outside Γ where
→ → →|) for all − →∈Γ x d(− x ) = min(|− x −− x I I
(5)
2. The volumes of the outer and inner compartments of the normal and control cases were calculated. The means (µO1 , µI1 and µO2 , µI2 ) as well as the standard deviations (σO1 , σI1 and σO2 , σI2 ) were calculated. 3. Three hypotheses test were performed to test the proposed concept, the first has a null hypothesis that µw1 > µw2 to compare the volumetric measures of the whole white matter. The second one has a null hypothesis µO1 > µO2 to test that the increment of volume (if proved by the first test) is resulting from the increment of the outer compartment volume. And, Finally, the third test has a null hypothesis that µI1 6= µI2 to test if the inner compartment played any role in the volumetric changes found between the different groups. 4. EXPERIMENTAL RESULTS The proposed algorithms were applied on both dyslexic and normal control brain images and this section presents sample results. Fig.3 shows the results of the brain extraction technique, Fig.4 shows a 3D volume of the segmented white matter for one of the dyslexic cases in the three different views: sagittal, planar and coronal.
is calculated. The boundary between the inner and outer compartments is obtained by detecting all the points that are far from the boundary by a distance d1 which can be mathemati− → cally expressed as X={→ x : d(− x )=d1}. In this sense, the set of points X represents a contour that is parallel to the outer contour and distant from it by d1 pixels. This concept can be illustrated in Fig 2.
(a)
The boundary between the inner and outer compartments represented by the intersection with the plane Z=1.5 3.5
Fig. 3. Results of skull stripping(a) Original MRI slice, (b) Skull stripped MRI slice
3 signed distance function
(b)
2.5 2 1.5 1 0.5 0 −0.5 1 1
0.5 0.5
0 The outer contour represented by the intersection with the plane Z=0 y
0 −0.5
−0.5 −1
−1
x
(a)
Fig. 2. Illustrating the idea of moving the outer contour using the
(b)
(c)
signed distance function, intersection with plane Z=0 represents the original contour and the intersection with plane z=d1 represents the new boundary
Fig. 4. Different views for the segmented white matter (a) Sagittal,
Having parcellated the white matter, the volumes are calculated and the hypothesis tests are performed, the results of all the previously described methodologies will be shown in the next section.
Fig.5 shows the results of the parcellation technique on a synthetic image for a circle, however fig. 6 shows the results of the parcellation technique when applied on segmented white matter slices.
3.3. Statistical analysis The brain extraction and segmentation represent the basic preliminary work that enables us to extract the white matter from each MRI stack, and the white matter parcellation is a basic step as well to get the region of interest. Having segmented and parcellated the white matter, the statistical analysis consists of the following steps; 1. The volumes of the whole white matter of all cases have been calculated. The means (µw1 and µw2 ) and standard deviation(σw1 and σw2 ) for the normal and dyslexic groups, respectively, were calculated.
(b) Coronal, and (c) Planar
(a)
(b)
(c)
Fig. 5. Parcellation results for a synthetic image for a circle (a) Original contour, (b) Singed distance map, and (c) The new boundary superimposed on the original image
The previous figures are meant to illustrate the efficiency of the algorithms used as preliminary work for getting the volumetric measures. Table1 shows sample results for the whole
(a)
(b)
(c)
Fig. 6. Parcellation results for an MRI slice (a) Original contour, (b) Singed distance map, and (c) The new boundary superimposed on the original MR image
white matter volume in normal control cases and dyslexic patients. Table 2 shows sample results for the volumes of the outer compartment at distance d=5 pixels Table 1. Sample Results for the volumetric measures of the whole white matter
Index
1
2
3
4
5
Normals Dyslexic
615.375 540.562
650.812 546.187
699.187 513.562
626.062 529.312
604.125 567.562
Table 2. Sample Results for the volumetric measures of outer compartment of the white matter
Index
1
2
3
4
5
Normals Dyslexic
469.494 447.555
516.435 446.786
584.037 419.589
517.446 441.196
510.297 447.663
Figure 7 shows a comparison between the volumetric measures of the normal cases and the dyslexic patients Comparison of the volumetric measures of the outer compartment
Comparison of the total white matter volume
700
800
Normal Control Cases Dyslexic 650
700
600
650
550 Volumes (ml)
Volumes ml
Normal cases Dyslexic 750
600
450
500
400
450
350
400
0
2
4
6
(a) Index
8
10
12
14
300
[1] V. B. Mountcastle. ”The Minicolumnar Organization of the Neocortex,” Brain, Vol. 120: 701-722, 1997. [2] D. P. Buxhoeveden and M. F. Casanova, ” The Minicolumn and Evolution of Brain: A Review,” Brain, Behavior and Evolution,Vol.375:1-25,2002. [3] K. R. Pugh, W. E. Menel and A. R. Jenner et al. ,”Functional Neuroimaging Studies of Reading and Reading Disability (Developmental Dyslexia),” MRDD Research Review, Vol.6, 2000 [4] P.Brambilla, A. Hardan and S.V.Nemi et al.,”Brain Anatomy and Development in Autism Review MRI Study,” Brain Research Bulletin,Vol.61,2003
500
550
of the inner compartments of the normal and dyslexic, has been rejected with 95% which emphasizes that the inner compartment did not affect the volumetric changes between the different groups and all the changes are resulting from the outer compartment which proves the hypothesis that has been postulated at the beginning of the paper: the change in the volumes is a result of the disturbance in the minicolumns; consequently, the connections between them through the white matter. For the future work, two directions are taken into considerations; first, analysis of the correlation between the minicolumns and other different parameters such as the gyrification index, gyral window using the deformable models will be considered. Second, diffusion tensor images are intended to be used to study the structural difference in the U fibers connecting the minicolumns to create a classifier that differentiate between two diseases. 6. REFERENCES
[5] S. Palmen, H. Engeland,P. Hof and C. Schmitz, ”Neuropathological findings in autism,” Brain,Vol.127, 2004 2
4
6
8
(b)
10
12
14
Index
Fig. 7. Comparison of volumes between the Normal cases and the Dyslexic (a) Whole white matter, (b) Outer Compartment
The results for the hypotheses tests explained in section 3.3 were very promising. The first hypothesis test is performed twice, once at confidence level α= 0.05 and another time at more restricted tolerance α =0.01 and for both times, the null hypothesis was accepted. The second test, as well, has been performed with several combinations. It has been done with the volumetric measures at different distances d1=5 and d1=7 ,as well as, different confidence levels α=0.05 and α=0.01 and the null hypothesis that µO1 > µO2 has been accepted in all the previous combinations. The third test has been performed with the similar combinations; the null hypothesis was accepted when d1=5 and α= 0.01, but it has been rejected in all the remaining combinations. 5. CONCLUSION AND DISCUSSION From the volumetric measures and the results of the hypotheses tests, it is very clear that there is a significant difference in the white matter volumes of the normal control cases and dyslexic patients. Also, it has been verified that the decrement of volume in dyslexia occurs in the outer compartment; since the hypothesis that: there is a difference between the volumetric measures
[6] S. Eliez, J.M.Rumsey, J. Giedd, J. Schmitt, A. Patwardhan and A. Reiss, ”Morphological Alteration of Temporal Lobe Gray Matter in Dyslexia: an MRI study,” J.Child Psychol.Psychat, Vol.41, No.6, June 1999 [7] M. Casanova, A. Switala, J. Trippe and A. Fobbs ”Minicolumnar Morphometry: a Report on the Brains of Yakovlev, Meyer, and Geschwind”Technicla Report, Department of Psychiatry and BehavioScience, University of Louisville, 2004 [8] N. El-Zehiry, M. Casanova, H. Hassan and A. Farag, ”Structural MRI Analysis of the Brains of Patients with Dyslexia” Proceedings of Computer Assisted Radiology and Surgery (CARS 2005), Berlin, Germany, June, 2005 [9] A. A. Farag and Hossam Hassan, “Adaptive Segmentation of Multi-modal 3D Data Using Robust Level Set Techniques“, in Proc International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI’04), Saint Malo, France, pp. 143-150, September, 2004. [10] R. Duda, P. Hart, and D. Stork. ”Pattern Classification ,” John Wiley and Sons Inc., 2001. [11] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces, Springer, 2003
