| dc.description.abstract | Among deep models in machine learning which have been very popular recently, Deep Belief Networks (DBNs) as a combination of stacks of Restricted Boltzmann Machines (RBMs) have shown excellent performance in signal fields, including dimensionality reduction, hierarchical representation, feature extractions, and classifications. On the other hand, Deep Neural Networks (DNNs) have indicated outstanding performance on feature learning tasks while deep Convolutional Neural Networks (CNNs) have been applied for classification tasks by reducing spectral variations and the correlations of the model which existed in images. However, to our knowledge, these approaches have not been extensively investigated on functional Near Infrared Spectroscopy (FNIRS) signals and Computed Tomography (CT) images. In this work, we attempted the applications of deep neural networks in those specific domains.
First, regarding to FNIRS signals, we presented an autoencoder DBN, which comprised six layers, could automatically extract relevant entropy features and construct multichannel FNIRS, including oxy-Hemoglobin (HbO), deoxy-Hemoglobin (HbR) and their differences over seven channels of Prefrontal Cortex (PC) regions. In addition, the classifier DBN by five layers, could classify FNIRS datasets, acquired from fifteen subjects. We also explored the usefulness of preprocessing steps of filters and Principle Component Analysis (PCA) before putting FNIRS datasets into our DBN models. The classification results between two classes, consisting of rest and task periods during the Stroop Task Experiments (STEs), were obtained with the best accuracy of 87.21 ± 4.02%, 83.16 ± 3.57%, 86.12 ± 3.87% of HbO, HbR and Diff channels, respectively. The neural signatures of stress levels between two discriminative types of hemoglobin did exist across individuals during our experiments.
Second, regarding to CT images, we implemented two hidden layers of the DNN model, using sparse autoencoders as unsupervised algorithm to obtain learned features, which allowed us to validate basic capabilities of the dataset. Besides, we trained a deep CNN which consisted of five main convolutional layers, followed by Rectified Linear Units (ReLUs) layers, max-pooling layers, three fully-connected layers and a final soft-max probability layer, to classify the high-resolution CT images into five anatomical organs in abdominal regions. As a result, we considerably achieved the classification accuracy of 83.74 ± 3.34 % in testing. We also visualized the layer representations on CT datasets, where they indicated the state-of-the-art performance, and could hold much promise to initialize further research on computer-aided diagnosis.
Keywords: functional Near Infrared Spectroscopy (FNIRS) signals, Prefrontal Cortex (PC), Stroop Task Experiment (STE), Human Mental Workload (HMW), Deep Belief Networks (DBNs), Deep Neural Network (DNN), sparse autoencoders, back-propagation, optimization, Convolutional Neural Network (CNN), convolution, ReLUs, pooling, Computed Tomography (CT). | en_US |
