Meeting Banner
Abstract #3256

Fully Connected Cascade Deep Architecture Neural Networks Outperform Support Vector Machines for Disease State Classification Using fMRI Data

Peng Wang1, Bogdan Wilamowski, Gopikrishna Deshpande2

1AU MRI Research Center, Department of Electrical and Computer Engineering, Auburn University, Auburn, AL, United States; 2AU MRI Research Center, Department of Electrical and Computer Engineering; Department of Psychology, Auburn University, Auburn, AL, United States

Brain disorder classification is traditionally done by Support Vector Machines (SVMs) due to SVMs capability of handling data of high dimensionality and superior training speed. SVMs are effective in correctly identifying non-ADHD subjects. However SVMs are ineffective in correctly identifying ADHD subjects. Two-stage Fully Connected Cascade Deep neural network architecture has been designed and modified experimentally. This FCC Deep NN architecture significantly excels traditional NN architecture, overcomes data unbalance issue, is capable of handling data of high dimensionality and easy to train, generates better results, and therefore outperforms SVMs in total.

Abstract PDF Presentation Video

Keywords

accuracies accuracy addition advantages architecture assuming attention auburn available back beats better bias biased brain broad capability capable cascade caused chose classes classification classifiers classifying clinical common competition complete component components computer connected consisted contest control convergence correction critical cross database datasets deep deficit derived diagnosis diagnostic dimensionality disease disorder done easy electrical engineering ensured error every excellent explained favored feature features feedback formed fully generalization generally give gives good hand head healthy illustrates inconvenient individually individuals inputs inside involving issues iteration latent layer learning limitations looking machine machines medicine motion networks neural noise noteworthy ones optimize outperform outperforms overcome overcomes patients performance possess potential principal process processed projects propagation proposed psychology ranked ranking reduced reduction releases reported represent respectively restricted root saved schematic separation series sets significance significantly since software square stage started steps studies subject subjects support taking train trained training trans unbalanced uniformly validation variables variance vector winning