Journal of Undergraduate Research
Volume 1, Issue 8 - May 2000

Can Experience Change the Brain? Attempts to Validate an Automated Computer Algorithm

Amy Richardson

INTRODUCTION

Brain Maturation

The relationship between brain maturation and experience is still largely undefined. Myelination, one of the major processes of brain maturation, involves the development of myelin sheaths, which increase the speed of neural transmission (Van der Knaap et al., 1991).

Visual inspection of MRI scans at different ages suggests that tissue composition changes primarily from gray matter to white matter and from gray matter to cerebrospinal fluid. The white matter of the brain consists of myelin sheaths and supporting glial cells. A previous quantitative study showed an increase in the overall volume of white matter well into adolescence (Paus et al., 1999).

MRI

Technological advancements in Magnetic Resonance Imaging have made new studies of brain maturation possible. MRI allows subjects to be studied in vivo, and thus, maturation of the same subjects over time can be examined. MRI scans also allow these maturational changes to be quantified.

Registration

Although previous studies have shown that the overall volume of the brain increases rapidly after birth and then levels off after two years, some post-mortem studies suggest that different anatomic regions of the brain mature at different rates. Registration programs allow images from the same person at different ages to be superimposed on one another and the difference image calculated. Thus, registration programs would allow brain development in specific anatomic regions to be quantified.

Segmentation Issues

Since the signal intensities of MRI scans show a continuous distribution throughout different tissue types, there is no known straightforward way to classify the types. Therefore, this study simply looks at the differences in signal intensity. The following equation for the signal intensity of the second scan was assumed:

St2 =St1+J+D

(where St2 = signal intensity of pixel in second scan
St1 = signal intensity of pixel in first scan
J = canner noise
D = developmental change)

 

Thus, when developmental change is positive, there is expected to be an increase in the proportion of white matter in the pixel. When the developmental change is negative, there is assumed to have been an increase in proportion of CSF in the pixel.

Visual Association Areas

Previous post-mortem studies have shown that although primary visual cortex increases in thickness until the 6th postnatal month, other cortical areas, such as the association areas, show a long and variable increase in cortical thickness that approaches maturity around 10 years from birth (Albert et al., 1999).

PURPOSE

The aim of this study was to measure the change in the amount of white matter occurring in normal children over a two-year period and to correlate these changes with behavior and age. A registration program was used to determine developmental increases in signal intensity from MRI scans. In this study, the changes in signal intensity of the visual association areas were ascertained.

The visual association areas in our subjects are hypothesized to have undergone developmental changes, specifically with an increase in the amount of white matter. These changes are hypothesized to correlate with the age, behaviors, and cognitive development of the child. The developmental changes are expected to decrease over age, as maturation processes in visual association cortex come to an end.

METHODS

Subjects

The population studied consisted originally of 31 subjects. These subjects were part of a longitudinal study of normal development. The children in the study had no neurological or psychiatric diagnoses and had a normal distribution of IQ. Each subject received two MRI scans. These MRI scans were taken an average of 2.22 ± 0.25 years apart. Data from two of the subjects was dropped after examining the images due to an unusually high amount of motion artifact. A third subject was removed due to an unusual amount of asymmetry in the measured area of the occipital lobes. The remaining 28 subjects consisted of 15 boys and 13 girls with an average age at first scan of 7.20 ± 1.14 years and an average age at second scan of 9.42 ± 1.26 years.

The MRI scans were transferred to a computer workstation, where they were displayed and analyzed with an image registration program written by Baba Vemuri, a professor in the Computer Science Department, and with programs written in PVWave.

Description of Registration Program

The image registration program was designed to measure developmental changes in brain structure and pixel intensity. The first and second MRI scans of each subject were registered with this program. To set the input parameters, the anterior commissure, visible from the midline sections of the MRI scans, was designated as the center of interest. During the registration of the two scans, the program performed a transformation from the first image to the second image using the designated center of interest. An example of the difference image computed by the registration program is shown in Figure 1. Figure 2 shows the difference image of one of the subjects who was removed from the study due to a large amount of motion artifact (Figure 2(a)).

Figure 1. Images from Registration Program for One of the Subjects in the Study.

Figure 1. Images from Registration Program for One of the Subjects in the Study.

(a) The subject's first scan (at time 1) on slice 56.
(b) The subject's second scan (at time 2) on slice 56.
(c) The overall difference images as computed by algorithm on slice 56,
(d) The positive difference image as computed by algorithm on slice 56.



Figure 2. Images from Registration Program for One of the Subjects Removed due to Motion Artifact.
(a) The subject's first scan (at time 1) on slice 54.
(b) The subject's second scan (at time 2) on slice 54.

(c) The overall difference images as computed by algorithm on slice 54.
(d) The positive difference image as computed by algorithm on slice 54.

Measurements

Image analysis was done with programs written in PVWave. Measurements of the occipital lobes were made between 20 and 40 Talairach mm lateral to the midline in both hemispheres. This region of the occipital lobe was chosen to obtain primarily visual association cortex. The occipital lobe was defined as the area bounded anteriorly by the parieto-occipital fissure.

RESULTS

A large portion of this project included improving the algorithm by solving technical problems that arose during attempts to automatically register the scans. Thus, the following processing techniques were developed:

A. Midline Section Alignment
B. The Negative Image Problem
C. Measurement Program­Visual Association Area
D. Compensations for Scanner Noise

Midline Section Alignment

The registration program was unable to perform the transformation accurately when the first and second scans' midline sections were on different slices. An example of this problem can be seen in Figure 3. A program was written in PVWave to set the midlines of the two scans for each of the subjects at the same slice number. All of the subjects' scans were run through this program before registration.

Figure 3. Images from Registration Program Showing the Midline Alignment Problem.

Figure 3. Images from Registration Program Showing the Midline Alignment Problem.
(a) The subject's first scan (at time 1) on slice 82.
(b) The subject's second scan (at time 2) on slice 82.

(c) The overall difference images as computed by algorithm on slice 82. Note the outlines of the subject's head from both time 1 and time 2.
(d) The positive difference image as computed by algorithm on slice 82.

The Negative Image Problem

The difference in the first and second scans of each of the subjects was originally coded in bytes. Since half of the pixels were showing a decrease in pixel value, these differences were negative and could not be expressed in bytes. Thus, the program could not differentiate changes in the direction that would be expected if gray matter changed to CSF from changes in the direction that would be expected if gray matter changed to white matter. To solve this problem, the registration program was changed to produce both a positive difference image and a negative difference image. The positive image shows developmental increases in signal intensity from the first scan to the second scan, which indicates that gray matter changed to white matter. The negative image shows decreases in pixel intensity, indicating that gray matter changed to CSF.

Measurement Program-Visual Association Area

The measurements of the occipital lobes initially showed poor reliability. A program was written in PVWave to standardize the area measured. In order to minimize the use of subjectively assessed boundaries, specific landmarks were used. The superior boundary of the occipital lobe measurements was defined in a medial section by placing a cursor at the superior portion of the parieto-occipital fissure. The anterior boundary was then defined in the same medial slice by placing a cursor at the posterior end of the splenium. These landmarks were projected onto a more lateral section, ± 20 Talaraich mm. For each of the sections from 20 to 40 Talairach mm, the area of the occipital lobes was measured by tracing the cursor from the marked superior landmark to the anterior landmark and then to the inferior pial boundaries. Measurements were taken by two different raters in order to ensure reliability using this method. This program overlaid the measurements from the occipital lobes of the second scan on the positive and negative difference images. For each slice, the area measured and a histogram of the signal intensity differences were stored in a computer file.

Compensations for Scanner Noise

After registering and measuring 16 subjects' scans, it was discovered that the registration program did not compensate for arbitrary differences in signal intensities due to variations in scanner settings. A post-processing technique was developed to compensate for these differences. The reversed negative histograms and the positive histograms were imported into a spreadsheet and the peaks of each of the histograms were determined. The peaks of the histograms were assumed to be at zero differences in pixel intensity if scanner parameters had been equivalent for the two scans. The histograms were then shifted accordingly so that all of the peaks lined up at zero. An example of an arbitrarily shifted histogram is shown in Figure 4.

Figure 3. Images from Registration Program Showing the Midline Alignment Problem.
Figure 4. Example of a shifted histogram.

The effects of scanner characteristic differences and developmental differences were simulated. This simulation suggested that the differences at the peaks were due to scanner noise and the differences near the edges of the distribution were due to developmental differences.

Data Analysis

The hypothesis was then tested. No simple relationship between the developmental changes and the age and/or behavior of the subjects was found. However, a pattern between the graphs of the proportion of pixels versus the developmental difference in signal intensity was found.

With large populations, a good measure of diversity is entropy. Entropy is the average amount of information required to select observations by categories (Krippendorff, 1986). The equation for entropy is:

H=-Sp(log2p)

The subjects' scans were divided into the following four classifications:

  1. Entropy high and roughly equal in the two hemispheres
  2. Entropy low and roughly equal in the two hemispheres
  3. Entropy lower for right hemisphere
  4. Entropy lower for left hemisphere

The histograms for each of the subjects were then averaged for each group. The averaged graphs are shown in Figure 5.


Figure 5. Graphs of the mean proportion of pixels versus developmental change in signal strength for each of the four classifications according to entropy.
(a) Entropy high and roughly equal in two hemispheres.

Subjects in Group: 7

Graph: flat with low peak

Interpretation: considerable environmental change
(b) Entropy low and roughly equal in two hemispheres.

Subjects in Group: 8

Graph: high, sharp peak

Interpretation: immaturity, not much change
(c) Entropy lower for the right hemisphere.

Subjects in Group: 8

Graph: higher peak for the right hemisphere

Interpretation: more changes in left hemisphere than right
(d) Entropy lower for the left hemisphere.

Subjects in Group: 5

Graph: higher peak for the left hemisphere

Interpretation: more changes in right hemisphere than left

The subjects were divided at the median entropy into high and low entropy groups. Using a t-test, the subjects with high entropy in the right hemisphere were found to have significantly better long-term memory ability at both time 1 (t=2.92;p<0.01) and time 2 (t=2.62,p<0.05).

DISCUSSION

Since the children who have a flatter graph of the proportion of pixels versus the developmental difference in signal also have higher long-term retrieval abilities, these children may be incorporating more information into their brain and, thus, making larger tissue rearrangements.

Additional research will be necessary to further validate the algorithm. More scans need to be processed and detailed observations should be recorded in order to ensure that the positive and negative differences reflect actual change in tissue characteristics. Various methods of validation could be used to test both the algorithm's method of registering the brains and our results that suggest brain changes and memory abilities are related.

As a result of this study, a new registration program was written by Jun Ye and Baba Vemuri of the Computer Science Department. This registration program is expected to pre-normalize the scans and, thus, to compensate for scanner characteristic differences. Further research is being directed towards analyzing the differences between the two methods of image registration.

REFERENCES

  1. Albert M., Diamond A., Fitch R., Neville H., Rapp P., Tallal P.. (1999) Cognitive
    Development. In: Fundamental Neuroscience (M. Zigmond, F. Bloom, S. Landis, J. Roberts L. Squire, Eds.), pp. 1313-1338. New York, Academic Press.
  2. Krippendorff, K. (1986). "Information Theory: Structural Models for Qualitative Data." Sage University Paper Series on Quantitative Applications in the Social Sciences, 07-062. Beverly Hills: Sage Pubns.
  3. Paus, T., Zijdenbos, A., Worsley, K., Collins, D.L., Blumenthal, J., Giedd, J.N., Rapoport, J.L., and Evans, A.C. (1999). Structural Maturation of Neural Pathways in Children and Adolescents: In Vivo Study. Science, 283, 1908-1911.
  4. Van der Knaap, M.S., Valk, J., Bakker, C.J., Schooneveld, M., Faber, J.A., Willemse, J., and Gooskens, R.H. (1991). Myelination as an Expression of the Functional Maturity of the Brain. Developmental Medicine and Child Neurology, 33, 849-857.

--top--

Back to the Journal of Undergraduate Research