Graph-Based Neural Image Analysis and Classification
Long, Samuel Seth
MetadataShow full item record
Since the first commercial magnetic resonance imaging machine was produced in 1980 by the Fonar company, the technology has become widespread. In 2007, the Center for Disease Control estimated there were 7,810 MRI machines in the United States. This pervasive technology has enabled accurate diagnosis of many conditions and provided data for research, but also presents challenges in data analysis. The availability of large neural imaging datasets has motivated the need for efficient methods to extract information such as structural correlations with given neural conditions. In this dissertation, we answer the question ``Can we achieve superior classification performance on 3D imagery by identifying discriminating features as paths in the oct-tree decomposition of the images?". Discovering structural correlations by hand can be a lengthy process, and seeking to automate this process requires an answer to the more general question ``What is the best method for discovering structural correlations in MRI data?". Graph representation of data has been shown to be effective in many problem domains, and we show that it is an effective means to represent neural MRI images for classification and correlation discovery. The Graph Neural Analyzer (GNA) is intended determine to what extent classification by identifying discriminating features as paths in the oct-tree decomposition of neural images is possible, how well the approach performs on a variety of classifications, and whether or not the technique is general enough to operate on more than one type of neural image. The human brain is highly variable between individuals, and structural changes related to a particular condition may not be obvious, particularly when the condition develops slowly and is thus a difficult candidate for a longitudinal study. Diagnosis of neural conditions often incorporates neural imaging, but many conditions are not known to be visible in structural MR images. Other conditions can only be diagnosed by significant radiation exposure from methods such as positron emission tomography. GNA operates by recursively subdividing the brain into octants to form an 8-way tree, where branches are subdivided only if the part of the image they represent contains multiple tissue types. The resulting trees are classified by discovering discriminating branches, producing a feature vector based on the discriminating branches, and classifying the feature vectors using a Support Vector Machine (SVM). Classification results are given for class distinctions such as age, education, Alzheimer's Disease, socioeconomic status, ethnicity, and Parkinson's Disease. In order to answer the question ``Is this method general enough to operate on more than one type of neural image", GNA is used to classify Diffusion Tensor Images. In order to determine if integration of multiple image types or processing methods can improve classification accuracy, merging of multiple trees and ensemble processing are evaluated. Orientation normalization is performed using the 3D tree representation, demonstrating that when using a 3D tree representation most images are most similar when oriented in the same manner. Alzheimer's Disease is found to strongly affect ventricle shape, and education is found to correlate with the shape of the medial longitudinal fissure and the Sylvian fissure, which may be due to increases in overall brain mass due to education. Gender is predicted primarily by information in the MRI regarding facial structure and head shape, however using diffusion tensor images, which primarily show white matter structure, it is classified correctly 68.0\% of the time. Age is found to be easier to classify than any of the above distinctions. The classifier is found to have 90.9\% accuracy differentiating scans of individuals 40 and younger from those from individuals 60 or older. Ethnicity is shown to affect brain structure such that classification on this distinction is correct 86.0\% of the time. Socioeconomic status may have to do with societal factors, but can also be classified correctly 66\% of the time. Image source (machine and location) strongly affects images, with classification accuracy over 95\%. Classification based on diffusion tensor images is less accurate in our tests than classification based on structural MR images, however education, age, image source, and gender are all shown to be differentiated by DTI white matter imaging. Ensemble classification reduces computational time required for the approach, but does not improve accuracy in our tests. Finally, discovered correlations are shown to be consistent with known correlations.