Deep learning of human brain activation

• Main Author: Chan, Yi Hao
• Other Authors: Jagath C. Rajapakse
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Computer science and engineering
• Online Access: http://hdl.handle.net/10356/77363
• Institution: Nanyang Technological University

It has been shown that deep neural networks can be trained to perform classification tasks on neuroimaging data. However, past studies typically use the entire set of features to train without knowing whether they are relevant for the classification task. This poses a problem as there are often many more features than number of instances in neuroscience datasets, resulting in a huge amount of parameters to train and a high chance of overfitting. Recent works on neural network interpretability have made it possible to identify salient features that are most influential in performing the classification. Using a deep feedforward neural network (DNN) to train on functional magnetic resonance imaging (fMRI) scans obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the DNN was able to differentiate cognitively normal (CN) subjects from patients suffering from Mild Cognitive Impairment (MCI) and patients suffering from Alzheimer's Disease (AD). Importance scores for the input features were computed and an empirically derived subset of features was obtained and to train another classifier. Using at most 25% of the features, this new classifier is able to give a better performance than the original classifier that used all the features. Also, we report brain connections and regions that the method found to be implicated in different patient groups. Besides finding a close correspondence with previous studies, it was also found that the uncus and medial temporal lobe are common distinctive regions for CN vs MCI and CN vs AD. It is proposed that this methodology can be applied to neuroimaging data across different imaging modalities and other neurological and neuropsychiatric disorders. Furthermore, knowledge of such characteristic features of a brain state can help narrow down the focus of clinicians and help design diagnostic systems with higher accuracy.