Executive Summary : | Functional MRI (fMRI) has gained popularity in recent years for various applications, including disease diagnosis, presurgical localization of brain regions, and cognitive trait prediction. However, the acquisition process is complex, time-consuming, and costly, especially for noncompliant subjects like infants, older people, and people with disabilities. To broaden the clinical applicability of fMRI, this project proposes a prediction model that infers information from task-fMRI data by predicting task activation maps using resting-state fMRI (rs-fMRI) and structural MRI, without actual task-fMRI data. The proposed graph signal regression-based approach models the input rs-fMRI data as a graph signal, treating each voxel as a node and defining functional connectivity values as edge weights between them. The target variable is a graph signal, whose node signal values form the required task activation map. This approach exploits additional relational information among different voxels, unlike conventional Euclidian approaches. A graph signal regression model consists of CNN and a graph convolutional network, but defining a common input graph for different subjects is a challenge. Different graph neural network architectures must be explored to design an optimal model for activation map prediction. A voxel-based regression approach can also be validated using conventional regression tools like LASSO or elastic net, or advanced deep learning-based methods like LSTM and RNN.
The obtained activation maps can be used to predict cognitive traits by learning their relationship using conventional learning models or the proposed graph signal-based regression model. |