Cerebral Glioma Grading Using Bayesian Network With Features Extracted From Multi-Modal MRI

Brain tumors are one of the most common types of central nervous system diseases threatening human lives and the preoperative grading of them is an important but difficult task. In recent years, more advanced MRI techniques (perfusion MRI and MR spectroscopic imaging) have brought more useful diagnostic information about tumors. Moreover, Bayesian Network is a powerful tool in the field of artificial intelligence with the ability of combining different kinds of information and probabilistic inference under uncertainty. Hence, we explored the diagnostic value of Bayesian Network in the application of cerebral glioma grading.In feature extraction, we used the following features which are all associated with tumor grade: mid-line displacement, perfusion MRI features, contrast enhanced T1W feature and MR spectroscopy feature. Specifically, we proposed a relatively simple method to measure mid-line displacement and we defined the contrast enhanced T1W feature based on clinical diagnostic criteria. For perfusion MRI, we chose the ratios of perfusion data from tumoral regions over those from contralateral normal regions as the perfusion features. For MR spectroscopic imaging, we proposed multiple search criteria based on quantitation from LCModel and implemented the visualization of metabolite alterations.In constructing Bayesian Networks for grading, we in turn used mid-line displacement, perfusion features, perfusion features with contrast enhanced T1W feature, MR spectroscopic features to construct Bayesian Networks. Network structures were determined using K2 algorithm, distribution parameters were learned using maximum likelihood estimation, and grading accuracies were computed using leave-one-out analysis. We found that in perfusion features regional cerebral blood volume, mean transit time and regional cerebral blood flow are most important for grading, and the grading accuracy can increase when they combine with contrast enhanced T1W feature. In MR spectroscopy features, we found that Cho/Cr, NAA/Cr and Lip13/Cr are most beneficial. We also explored Bayesian Networks combining different kinds of features with parameters learned using maximum likelihood estimation and expectation-maximization. We found that perfusion features and contrast enhanced T1W feature are most beneficial, MR spectroscopy features come the second and the last is mid-line displacement. We also found that expectation-maximization can give better parameter estimation than maximum likelihood estimation but at the cost of more time.