# ROI-Based Statistical Analysis, Representation, and Discovery of Human Brain Functions

Development Team:
Faculty:    Nora Volkow (BNL Medical Group)
Klaus Mueller (Computer Science)
Wei Zhu (Applied Mathematics and Statistics)
Students:  Tom Welsh and Jeffrey Meade (CS)
Juan Li, Radha Panini, and Shurou (Sue) Wu (AMS)

Papers:

Online Story:

Introduction

The overarching goal of this project is to investigate and discover correlational relationships of different regions of the brain. Evidence for these relationships are gathered by PET-imaging a number of human subjects, both under baseline conditions and under the influence of a drug, in our case Ativan. The correlations in the brain activity are calculated on the basis of predefined anatomical regions-of-interest (ROIs), for now modeled as spherical regions. The correlation coefficient is then employed to quantify similarity in response, for various regions during an experimental setting. To account for inter-human anatomical variability, each test subject's volumetric brain data is first transformed into a common anatomical coordinate system (e.g, Talairach-Tournoux). Statistical parameters that can be used to characterize various brain functions include:

• The correlation (including the Pearson product-moment correlations, the partial correlations and the canonical correlations) matrix..
• The ROI clusters from the cluster analysis.
• The principal components and the factor analysis output. (The PCA and the FA are similar in their function and output, but different in their assumptions and derivations).
• Differential relationships such as the difference of two correlation matrices (to view the change of functional relationships).
• The times series.
The amount of statistical data can be enormous, and effective tools are essential for the brain researcher to grasp and discover functional relationships quickly from the statistical data. BrainMiner is a visualization tool that facilitates this task.

Viewing in 2D

One approach to view statistical brain relationships is by overlaying the statistically significant voxels on top of a high resolution MRI and view the data as 2D slices, either in flip mode or side-by side:

Fig. 1: The circular ROIs are colored according to their correlation with respect to a root-ROI, marked by a red cross. The rainbow color scheme is used, where the color blue stands for highly negative correlation and the color red stands for highly positive correlation. Green and yellow stand for mildly negative and positive correlations, respectively. There are apparently no extremely strong correlations in this configuration.
The 2D approach works well as long as the axial dimension is not important. However, the decomposition of the dataset into 2D slices for visualizing 3D relationships becomes limiting when relationships are widely spread over the brain.

Viewing in 3D

To account for the problems with the 2D approach, we have also developed (in addition to the 2D viewer) a 3D visualization interface that displays correlational data for each ROI along with an MRI volume and a digitized version of the Talairach atlas. Both can be sliced in 3 orthogonal directions and can be overlaid on top of each other. Here is a screenshot of the Graphical User Interface (GUI) of the system, where a basic view with a few ROIs is shown:

Fig. 2: Here we see the Graphical User Interface (GUI) of our newly developed 3D brain visualization software, along with a basic view of a small number of ROIs embedded into a cut-out area of a normalized/standardized MRI brain. For now, all ROIs are spherical in shape. Similar to the 2D viewer, the colors of the ROIs denote the strength of the correlational relationship, on a rainbow scale. The root ROI is colored in yellow. The GUI allows the user to slide the cutting planes up and down and back and forth, to rotate the volume, and to select certain brain surfaces, such as white matter, grey matter, and skull to be semi-transparently superimposed. The correlation thresholds can also be selected, and many more features are available.
The number of ROIs to be displayed, however, can become quite large (about 120-140), which poses challenging problems in the visualization task: In a space too crowded with statistically significant  ROIs, it becomes very hard, if not impossible, for the user to tell the 3D positions of the individual ROIs. To overcome these difficulties, a number of techniques were investigated:
• Superimposing a Talairach atlas slice that can be slid up and down the volume:
Fig 3: A movable sheet that shows the slice of the Talairach atlas at the specified height. ROIs that intersect the sheet are highlighted by a ring.
• A single light source placed above the volume in a fixed position, providing specular lighting cues for the height and depth of each ROI sphere (this can be seen in Figs. 2 and 3).
• Enhancing the ROIs by colored halos, where the colors code their height and depth on a rainbow color scheme. The ROIs are connected by iso-lines to the MRI volume cuts which suggests their position in 3D space:
Fig. 4: ROI halos, painted in colors corresponding to ROI height and depth (the rainbow color scheme is used). Dashed iso-height and iso-depth lines emanate from the ROIs and pierce the MRI volume slices at the ROI depth and height.
• Projecting a colored grid onto the volume cuts, again encoding height and depth on a rainbow map. Colored shadows cast onto the exposed volume slices provide additional cues:
Fig. 5: Iso-lines, with height coded into rainbow colors, are drawn onto two of the three orthogonal MRI volume cuts. The ROI halos are coded in height using the same color scheme. The ROI position with respect to the third MRI cut is suggested by shadows cast by the ROI spheres onto that plane.
• Projecting the ROIS onto the brain iso-surface, such as white or grey matter, or skin:
Fig. 6: Since the ROIs are mostly located close to the brain surface, i.e. on the brain cortex, one can generate a comprehensive, EEG-like, view by projecting the ROIs onto the cortex surface and paint the projection in the correlation color.
• Grouping ROI networks into composite polygonal objects, which reduces the object complexity of the scene:
Fig. 7: For now we simply increased the radius of the ROI spheres until they just touched. This approach is rather effective. For the future we plan to estimate the actual hull of a set of ROIs of similar brain function and display this hull as a polymesh.
Results
To test the software and develop its capabilities, PET fluoro-18-deoxyglucose (FDG) images were analyzed and displayed for two major drug addiction studies. The first study included 30 subjects the second included 40 subjects both under baseline and drug conditions. Metabolic activity was measured for each subject as the average intensity signal for a given ROI defined manually by a trained medical doctor. ROI locations for the first study included 424 anatomically significant regions while the second study included 120 regions. Correlation matrices were generated for each of these ROI datasets. Three major statistical measures were subsequently generated given the correlation matrix for each study: 1) Principle Component Analysis, 2) ROI cluster analysis and 3) Factor Analysis. Visualization software was compiled using libraries for OpenGL, the Fast Light Toolkit (FLTK) and written in C++. Versions were compiled for Irix 6.4, Windows NT 4.0 and Linux and run on an SGI O2/R1000, a Pentium II class laptop and a Pentium 233 respectively. The visualization interface displayed a 3D MRI volume representative of Talairach coordinates which was sliced as three half planes of axial, coronal and saggital views. The slices were drawn using 2D texture mapping. The user could define the slice location along these three axis by dragging each slice with a mouse button. The surface of the brain was also displayed as a polygonal mesh generated from the Marching Cubes algorithm. ROIs were drawn as spheres but were obscured by the surface depending on the current slice locations. Correlation value was represented for each ROI as a color intensity in relation to a selected "root" ROI. A side window also displayed the current slice as either the Talairach atlas digitized or the MRI volume along with a 2D representation of the ROIs (circles). In addition, all objects could be rotated together with the mouse (trackball interface) to provide a viewpoint from any direction.

Conclusions

BrainMiner allows interactive exploration of both the 120x120 and 424x424 correlational data on many platforms. It is cuurently in use by a number of brain researchers and is being refined almost on an hourly basis.