Ensemble Tree Learning Techniques for Magnetic Resonance Image Analysis

This paper shows a comparative study of boosting and bagging algorithms for magnetic resonance image (MRI) analysis and classification and the early detection of Alzheimer's disease (AD). The methods evaluated are based on a feature extraction process estimating first-order statistics from gray matter (GM) segmented MRI for a number of subcortical structures, and a learning process of an ensemble of decision trees. Several experiments were conducted in order to compare the performance of the generalization ability of the ensemble learning algorithms for different complexity classification tasks. The generalization error converges to a limit as the number of trees in the ensemble becomes large for boosting and bagging. It depends on the strength of the individual trees in the forest and the correlation between them. Bagging outperforms boosting algorithms in terms of classification error and convergence rate. The improvement of bagging over boosting techniques increases with the complexity of the classification task. Thus, bagging is better suited for discrimination of mild cognitive impairment (MCI) from healthy controls or AD subjects than boosting techniques.