Automatic detection of Attention Deficit Hyperactivity Disorder (ADHD) using resting-state - functional Magnetic Resonance Imaging (rs-fMRI) data is challenging due to the small number of samples, variations in acquisition technologies/techniques, imbalance in class distribution and high dimensionality of the data. In this paper, a Convolutional Neural Networks (CNNs) based classifier is proposed to automatically diagnose ADHD. CNN is a complex deep learning method with a wide range of applications in hand-writing recognition, computer vision, image restoration and medical image analysis. CNN architectures are usually designed based on empirical observations or are derived from existing popular networks in literature. Determination of optimal architecture is imperative for achieving the best classification performance in deep networks. This paper proposes a Deep Network Optimizer (DNO), a data-driven method to modify the number of fully connected layers and/or nodes to determine the optimal CNN architecture. DNO analyses the network bias, variance and training performance to determine the network significance. The network significance is used to derive threshold-free growing and pruning strategies to evolve the CNN network. The results show that CNN architecture evolved using the DNO achieves state of the art accuracy of 80.39% on the ADHD200 dataset.
Paper
Full text
Deep Network Optimization for rs-fMRI Classification
Semantic Scholar · Computer Science · 2019
Abstract
Automatic detection of Attention Deficit Hyperactivity Disorder (ADHD) using resting-state - functional Magnetic Resonance Imaging (rs-fMRI) data is challenging due to the small number of samples, variations in acquisition technologies/techniques, imbalance in class distribution and high dimensionality of the data. In this paper, a Convolutional Neural Networks (CNNs) based classifier is proposed to automatically diagnose ADHD. CNN is a complex deep learning method with a wide range of applications in hand-writing recognition, computer vision, image restoration and medical image analysis. CNN architectures are usually designed based on empirical observations or are derived from existing popular networks in literature. Determination of optimal architecture is imperative for achieving the best classification performance in deep networks. This paper proposes a Deep Network Optimizer (DNO), a data-driven method to modify the number of fully connected layers and/or nodes to determine the optimal CNN architecture. DNO analyses the network bias, variance and training performance to determine the network significance. The network significance is used to derive threshold-free growing and pruning strategies to evolve the CNN network. The results show that CNN architecture evolved using the DNO achieves state of the art accuracy of 80.39% on the ADHD200 dataset.