Functional magnetic resonance imaging (fMRI) is the dominating approach to research in the mapping of neural activity in the human brain. State of the art data analysis techniques employ a statistical parametric mapping (SPM) strategy to convert raw signal into interpretable images by processing data in a pipeline of task-specific modules. This approach, despite its simplicity and reliability, presents a set of inconveniences, including low interconnectivity among modules, resulting in suboptimal solutions. The main problem is that each individual module within the pipeline works in isolation; that is, the required parameters are (often manually) chosen by an expert without any feedback about the implications of this choice to the subsequent modules.
The goal of the action is to make a major contribution to the field by replacing the step-by-step data processing pipeline by a deep neural network. We hypothesise that this new approach will achieve better solutions by propagating the effects of module-based decisions through the network, jointly optimising the whole processing pipeline. We aim at establishing a new paradigm that greatly improves optimisation and interpretability of state of the art data processing pipelines for fMRI.
By approaching the problem in the way it is explained in the previous paragraph, we are also tackling a set of other subgoals, as the use of multivariate techniques to account for interdependencies among features, as well as the ability to capture existing non-linear relationships in the data. Moreover, the proposed architecture can be used to analyse the resulting data processing pipelines, a tool that helps in better understanding the influence that each module has to the final task being investigated.
It is also important to assess the complexity of transferring these networks to other data acquisition modalities, such as electroencephalography (EEG), where temporal resolution plays an important role, and evaluate the technology that best adapts to these constraints. In this respect, we developed a system using cutting-edge computer vision techniques to automatically score and interpret sleep stages by means of EEG data, showing promising results towards its applicability in a neurofeedback environment, where subjects learn relaxation strategies guided by EEG technology.
Success in tackling the aforementioned goals has a great impact in improving the quality of data processing pipelines for fMRI and EEG; by providing a set of improved tools for analysing neurological data; which ultimately will aid in better understanding the brain functioning for both healthy and pathologic brains, and even help in the treatment of certain diseases.
In the current project, we have demonstrated the feasibility of achieving the proposed goals and released a freely accessible easy-to-use GPU-powered framework to perform automatic multivariate non-linear data analysis of neurological signals. We think that these achievements will contribute to paving the path for other research projects to adopt a similar approach and take advantage of the rapidly growing deep neural network field to be applied in those areas where advanced data processing of neurological measurements is required.
