Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data

The overall goal of this procedure is to determine signature Covance pattern imaging biomarkers of neuro functional brain disorders. This is accomplished by first obtaining single volume brain scans of a set of clinically pre-diagnosed patients and an age and preferably gender matched set of normal controls using modalities such as PET spect, A-S-L-F-M-R-I, or V-B-M-M-R-I. The second step is to spatially normalize the images to a common stereotactic template.

Next, a multivariate technique. The Scaled SubProfile model of principle component analysis or S-S-M-P-C-A is performed and the output principle components or PCs and their corresponding subject scores are examined. The final step is to determine the best PC or combination of PCs that define a Covance pattern that discriminates patients from controls.

Ultimately, the Covance pattern may be used to score individual prospective subjects on their expression of the disease. Advantage of this method relative to some of the popularized single voxel comparative techniques is the ability to identify and measure the activity of brain networks. These are distributed regional Covance patterns in the brain.

Most importantly, it allows one to measure the activity of a given network in the brain and to summarize it by a single number that is a major attribute for clinical investigation. This method can help answer key questions in the field of neuroscience, such as identifying the metabolic networks for different neurological disease that present similar symptoms in the early stages of disease progression. The implications of this technique extends towards both differential diagnosis and treatment planning for diseases such as Parkinson's disease because a single subject score can be obtained, which can then be applied to differential diagnosis and treatment planning.

Though this method can provide the insight into Parkinsonian syndromes. It can also be applied to other neurodegenerative disorders such as hunting's disease and Alzheimer's disease. Generally, individuals new to this method will struggle with preconceived univa concepts that consider the effects of disease to be localized in isolated regional clusters.

Dr.James Moler first had the idea for this method when he applied this concept to study dementia in AIDS patients. Visual demonstration of this method is important for comprehending the processes involved in network analysis. However, the actual steps involved in using our programs are not difficult to learn because they're highly automated To study brain metabolism using positron emission tomography or PET first administer a radionuclide tracer such as 18 F fluoro, deoxy glucose to each patient in the disease group following a fast of at least 12 hours off medications and to each normal control.

Allow 45 minutes for the tracer to reach equilibrium in the tissue. Then scan each individual while they're at rest with eyes open for pattern derivation. Scan an equal number of gender and H matched patients and controls once the scan is complete.

Transfer image data to a workstation and convert it to an appropriate format for analysis. The windows-based MATLAB analysis software used here requires conversion from ge advanced formats to analyze or nifty format images. Next, normalize each subject's image to a common stereotaxic space using a standard neuroimaging software package such as SPM, so that there is a one-to-one correspondence of voxel values between subjects.

Then apply a gray matter mask to the normalized image in order to limit the analysis to gray matter areas. If an external mask is not available, create individual subject masks by excluding values lower than a specified threshold. Percent of each subject's maximum value and multiply the masks to define a composite gray brain matter voxel space for the analysis.

In order to perform multi-variate scaled SubProfile model principle component analysis or S-S-M-P-C-A, use the in-house routines scan VP available on the Feinstein neuroscience website. Also available as a statistical parametric mapping toolbox, SSM PCA on the scan VP menu, select statistics. Then SSM, then voxel based, and then PCA.

Next, check select controls on the menu and pick the previously normalized control images. Then select image files to enter the normalized patient images. Select an appropriate mask or threshold and check the other options on the menu.

Next press process and enter a name for the experimental output. Wait for the display to view principle component images of interest in their corresponding score files for other software. Emulate the algorithm executed by the automated process just described.

In order to accomplish this first mask data using an available zero one image mask to remove unwanted areas of the voxel space such as white matter and ventricles as described in the pre-processing step. Next, convert each subject's 3D masked image foxhole data to one continuous row vector by appending sequential scan lines from consecutive planes form group data matrix so that each subject's data corresponds to a specific row of the matrix. Each column then represents a particular voxel across subjects.

Once each subject has been loaded into the matrix, transform each data entry to logarithmic form. Then center the data matrix by subtracting each row average or subject mean from the row elements. The average of all centered rows represents a characteristic group mean log image, term to group, mean profile or GMP.

Next, subtract the column means from the corresponding matrix column elements. Each row of the double centered matrix represents a residual image termed a subject residual profile or SRP whose elements represent deviations from both the subject S and voxel V group means this occurs according to the equation shown here where S-R-P-S-V is an element of the SRP matrix corresponding to subject S and voxel. VDSV is the original data value mean S is the subject mean value and GMPV is the group mean value of voxel V.Next, construct the subject by subject Covance matrix C of the composite double centered SRP data matrix by computing the non normalized Covance between each pair of rows of the subject residual profile matrix.

Using this equation shown here where CIJ is an element of the symmetric Covance, matrix C and c, RRP IV and CRP JV are corresponding voxel elements of the CRP matrix rose INJ that are multiplied and summed over all voxels. Then apply principle component analysis to the subject by subject Covance matrix C.The results will be a set of subject score egen vectors with associated egen values. Next weight each vector by multiplying by the square root of its corresponding egen value.

The set of score eigen vectors is represented by the columns of the matrix S shown here. Next, determine voxel egen vectors for the same set of egen values by multiplying the score vector matrix by the transpose of the subject residual profile matrix. This creates an array P of voxel pattern, egen vectors in a descending order of egen values according to the equation shown here where P is the matrix of principle component voxel egen vectors.

SRPT is the transpose of the SRP matrix and S is the score vector matrix then transform each vector to a principle component image. Each principle component image is attributed to a percent of the total variance accounted for or VAF corresponding to the relative size of its igon value. Examine the results of the principle component analysis to determine patterns that are associated with high accounted for values.

Voxel pattern vectors and scores are Z transformed so that their values represent positive and negative standard deviations from mean values. Next, attempt to identify regional deviations associated with the disease that is being studied. Then examine scatter plots of the scores corresponding to each principle component pattern.

An optional receiver operating characteristic plot can also be generated to identify a disease specific pattern. Evaluate the differentiation of subject scores between patients and controls by examining the P values of the corresponding two sample T tests and area under the curve values. Next vector normalize and linearly combine the selected principle components to yield a single disease related pattern.

To accomplish this, the MATLAB function, GLM fit is used to determine coefficients based on logistic or other regression models apply to subject scores. Once a biomarker pattern has been identified, you can evaluate its score expression in a prospective subject from that individual's scan, this operation is performed by TPR on the menu. To accomplish this use a simple computation of the internal dot vector product of the subject's SRP vector and the GIS pattern vector.

Finally, further validate the resulting potential pattern biomarker by bootstrap resampling using external software and by forward validation of independent subject groups. Forward validation is facilitated by using TPR to determine scores for a prospective cohort. Shown here are axial displays through the origin of the first four principle component images of the SSM analysis of 10 clinically diagnosed Parkinson's disease patients with 10 age and sex matched controls.

The hot colors indicate relative increases in metabolic activity. Within that principle, component's contribution to the overall subject residual profile, whereas cold colors indicate relative metabolic decreases. Principle component one had the largest variance accounted for or VAF value and the smallest P value representing high significance and is the only one that could be considered alone a potential independent biomarker.

The significant regions of variation that contribute to the overall pattern are evident when displayed in orthogonal views Over an MRI structural image background, the combination of principle components can be used to try and increase significance such as with the case of components one and four, where when combined lowers the P value suggesting higher significance. These graphs show the independent bar graphs and scatterplots of the derivation subject Z scores of principle component one and principle component four only. The first principle component significantly discriminates patients from controls.

While the fourth demonstrates a trend. The linear logistic combination of principle component one and principle component four improves discrimination. As previously indicated, the combined pattern demonstrates perfect separation at a Z-score threshold of 0.9.

Combining principle component three shows a slightly increased discrimination over the combination of one and four, but all four principle components did not improve data discrimination due to the non-discrimination capacity of principle.Component. Two, prospective score evaluation of patterns for principle component one and combined patterns of principle components one and four principle components one, three, and four and principle components 1, 2, 3, and four for 22 normal controls and 22 PD patients are shown here. The area under the curve values and sensitivity appear to decrease for more than two combined principle components.

While P values tend to become less robust, an insignificant improvement was noted for PC one four over PC one alone in a UC and P values. In contrast to the significant difference predicted in the derivation sample. Once scans from patients with a disease and control subjects are required the identification and validation of realistic and valid disease patterns, it becomes a straightforward procedure using this algorithm After the disease patterns have been identified and validated, other methods like correlation analysis of patient scores with independent clinical measures can be used to answer additional questions.

For example, whether the imaging biomarker reflects motor or or cognitive manifestations of disease or whether the expressions of disease patterns can help differentiate between different conditions that present similar symptoms in hard to diagnose clinical cases. Don't forget this procedure cannot be directly applied to time series functional MRI data because of the dynamic fluctuations in response and low signal to noise ratio associated with FMRI signals.