A novel ensemble approach on regionalized neural networks for brain disorder prediction

With the advance in neuraimaging technology, such as functional Magnetic Resonance Imaging (fMRI), more brain image data become available, and mining the brain images emerges as a hot area in data mining communities. In contrast to ordinary images extensively studied in computer vision communities, brain images can be parcellated into different Regions of Interest (ROIs) due to the underlying functional and structural organization of human brain. Distinct diseases are highly relevant to a specific set of ROIs. In this paper, in order to leverage regional patterns to predict diseases and identify relevant ROIs, we present a novel Ensemble Approach on Regionalized Neural Networks (EARNN). EARNN is of a two-stage learning architecture. Its first stage consists of training a set of Regionalized Neural Network models, each of which learns from a brain region by using a proposed stochastic pooling operation. Its second stage adopts an effective EM algorithm to ensemble a mixture of Regionalized Neural Networks and iteratively estimate the mixing coefficients. We test our model on two real fMRI datasets. Experiments demonstrate that by effectively exploiting regional patterns and structural information, EARNN achieves the superior performance in comparison to previous approaches. In addition, according to our domain knowledge, EARNN successfully identifies relevant ROIs of diseases.