MRI feature extraction using genetic algorithms

Abstract

Traditional machine vision techniques apply a feature extraction step before any classification, but this is not commonly done for magnetic resonance images. In this study the authors propose to discover optimal feature extractors for MRI to increase segmentation accuracy. Genetic algorithms are applied using a fitness function based on known class labels, and on a fitness function that can be applied to data without ground truth. Both fitness functions allow the discovery of good features, that can be applied outside the data that was used for the search. An increase in the tumor true positive rate for an MRI volume using fuzzy c-means (FCM) was found from 78.7% to 91.3% of all tumor pixels with constant false negative rate. This approach may lead to significantly improved MRI segmentation, which is needed in particular for multicenter trials for brain tumor treatment.