In this thesis, we propose computational network models for human brain. The models are estimated from the fMRI measurements, recorded while the subjects perform a set of cognitive tasks. In order to represent high level cognitive tasks of human brain, such as, complex problem solving, working memory, emotion, language, social activities, relational activities and gambling, by dynamic networks, we employ supervised and unsupervised machine learning techniques.
In the first part of this thesis, we propose an unsupervised multi-resolution brain network model by employing a deep learning architecture, called stacked denoising auto-encoder (SDAE). It is well-known that human brain operates at multiple frequency bands for different cognitive tasks. For this reason, we decompose the signal into multiple sub-bands using Wavelet transform and estimate a set of local meshes at each sub-band. This approach enables us to analyze the network properties of human brain at different frequency bands. Then, we use a deep learning architecture, called stacked denoising auto-encoders to learn low-dimensional connectivity patterns (features) from the constructed mesh networks. Finally, the learned connectivity patterns are concatenated across different frequency sub-bands and clustered using a hierarchical clustering method. Results clearly show that our proposed model successfully decodes the cognitive states of the Human Connectome Project (HCP) task dataset, yielding high rand index and adjusted rand index values.
In the second part of this thesis, we propose a supervised dynamic brain network model to decode the cognitive subtasks of complex problem solving. At first, the raw fMRI images are passed through a preprocessing pipeline that decreases their spatial resolution while increasing their temporal resolution. Then, dynamic functional brain networks are constructed using neural networks. The constructed networks successfully distinguish the phases of complex problem solving. Finally, we analyze the network properties of the constructed brain networks in order to identify potential hubs and clusters of densely connected anatomic regions during both planning and execution phases of complex problem solving task. Results clearly show that there are more potential hubs during planning and that clusters are more strongly connected in planning compared to execution.