Abstract
Localization of cortical regions of interests (ROIs) in structural neuroimaging data such as diffusion tensor imaging (DTI) and T1-weighted MRI images has significant importance in basic and clinical neurosciences. However, this problem is considerably challenging due to the lack of quantitative mapping between brain structure and function, which relies on the availability of multimodal training data including benchmark task-based functional MRI (fMRI) images and effective machine learning algorithms. This paper presents a novel joint modeling approach that learns predictive models of ROIs from concurrent task-based fMRI, DTI, and T1-weighted MRI datasets. In particular, the effective generalized multiple kernel learning (GMKL) algorithm and ROI coordinate principal component analysis (PCA) model are employed to infer the intrinsic relationships between anatomical T1-weighted MRI/connectional DTI features and task-based fMRI-derived functional ROIs. Then, these predictive models of cortical ROIs are evaluated by cross-validation studies, independent datasets, and reproducibility studies. Experimental results are promising. We envision that these predictive models can be potentially applied in many scenarios that have only DTI and/or T1-weighted MRI data, but without task-based fMRI data.