Optimized bilevel classifier for brain tumor type and grade discrimination usingevolutionary fuzzy computing

Abstract

In this paper, an optimized bilevel brain tumor diagnostic system for identifying the tumor type at the first level and grade of the identified tumor at the second level is proposed using genetic algorithm, decision tree, and fuzzy rule-based approach. The dataset is composed of axial MRI of brain tumor types and grades. From the images, various features such as first and second order statistical and textural features are extracted (26 features). In the first level, tumor type classification was done using decision tree constructed with all features. Further evolutionary computing using genetic algorithms (GA) was applied to select the optimal discriminating feature set (5 features) and classification using the decision tree constructed with the reduced feature set resulted in better performance. In the second level, grade classification, a fuzzy rule-based approach was used to resolve the uncertainty in discriminating the tumor grades II and III. Membership functions of all grades were defined for all features extracted from brain tumor grade images, to derive the fuzzy inference rules for grade discrimination. Similar to type classification with GA, better grade discrimination performance was exhibited with fuzzy inference rules derived using optimal feature set (13 features) using GA. Overall performance comparison of the proposed bilevel classifier with all features vs GA-based feature selection, shows that evolutionary computing combined with fuzzy rule-based approach is successful in reducing false positives, thereby enhancing classifier performance.