Machine learning in functional magnetic resonance neuroimaging analysis



Journal Title

Journal ISSN

Volume Title



Magnetic Resonance Imaging (MRI) is one of the most advanced non-invasive neuroimaging techniques available. Functional information is obtained using MRI by measuring the changes in local magnetic fields caused by the changes in the blood oxygenation concentration, which is an indirect measure of neuronal activity. The actual Blood Oxygenation Level Dependent (BOLD) changes are only about 2-5 % of the absolute signal intensity and are almost completely embedded at noise levels. Thus, sophisticated statistical and scientific tools are required to extract any useful information. However, achieving accurate and correct interpretation is still challenging because of low signal to noise ratio, high noise levels, artifacts, and lack of ground truth. Moreover, apart from structural MR imaging, modalities including functional, spectroscopy, and diffusion imaging are generally considered to be research tools with very little clinical application. Currently, many neurological disorders cannot be fully diagnosed until further progression of the disorders, making their treatment very difficult, sometimes impossible. Computational neuroimaging based on various modalities of MRI data offers great potential in providing quantitative biomarkers for such disorders. However, the complexity of most neuroimaging data makes analysis by conventional methods not computationally efficient. The recent development of deep learning algorithms for extracting relevant features from complex data sets suggests their application in generating more accurate neuroimaging-based biomarkers for early detection and diagnosis of neurological disorders. Synergy between deep learning models and functional MRI (fMRI) has potential to translate fMRI as a routine clinical technique but is limited by the complex fMRI data structure and its high susceptibility to noise and artifacts. The main aim of this study is to develop and investigate the application of preprocessing and analysis technique that are data driven, less susceptible to noise, and preserve the spatiotemporal content of fMRI data that can thus be used to train and implement a deep learning model to extract features and identify neurological disorders in a clinical environment. Moreover, a direct visualization technique allows quick preliminary interpretation and quality assessment of the fMRI data. First, a data-driven algorithm for signal drift and spontaneous fluctuation removal is developed. This algorithm makes use of Principal Component Analysis (PCA) to estimate temporal signal drift that is unique to each dataset. The main advantage of this algorithm over existing algorithms is that it is data-driven and does not require any predetermined models. Second, we developed a functional parcellation technique that produces functionally and spatially homogenous brain regions. In addition to resting state fMRI, it can be used for task fMRI data analysis. For task fMRI, this is a model-free technique to obtain information about whole-brain activation and hemodynamic response variability. Third, a temporal visualization algorithm was also investigated. This algorithm made use of the t-SNE technique and could not only be used to visualize the fMRI data but also be used in the identification of major changes in the brain states that correlate with the task paradigm. The correlation between brain state change and task paradigm can be used to assess the involvement of subjects during fMRI scans and act as a quality control measure. Apart from that, a MATLAB-based fMRI simulator was developed that can generate quasi-realistic 4D fMRI volumes as desired by the user. To generate realistic synthetic data, various fMRI noise sources and artifacts were studied and modeled. Such 4D fMRI-simulated data can serve to hypothesize ground truth for experimentally acquired data under both task-evoked and resting state designs in the investigation of localized or whole-brain activation and functional connectivity patterns. Finally, we propose to synergize deep learning with functional neuroimaging. Much information is present in the fMRI data that cannot be extracted directly using traditional techniques due to the complexity of data acquisition and data structure. Deep learning, on the other hand, has the potential to extract such information. A volumetric 3D Convolutional Neural Network (CNN) model was designed and trained that can extract spatial and temporal features and classify BOLD fMRI data for Alzheimer’s Disease (AD). Such classification can identify AD in the Mild Cognitive Impairment (MCI) stage, which can be useful in early detection of AD. Detection of dementia in an early stage can be applied to slow the progression of the disease and can also be used to establish fMRI-based biomarkers for dementia, thus taking a step toward the clinical application of functional MRI.



Functional magnetic resonance imaging (fMRI), Machine learning, Deep learning, Alzheimer's disease, Functional magnetic resonance imaging (fMRI) simulator, Resting state functional magnetic resonance imaging (fMRI) clustering, Functional magnetic resonance imaging (fMRI) visualization