Population learning of structural connectivity by white matter encoding and decoding

There is rapidly growing interest in analyzing brain connectivity at the population level, to detect factors that affect brain networks and fiber architecture. Even so, we still lack a fundamental model of brain fiber tracts and white matter (WM) connectivity, making it challenging to identify representative and diagnostically informative patterns. To bridge this gap, we introduce a framework to learn structural connectivity patterns from diffusion tensor images (DTI) of the brain. This novel strategy encodes key WM tracts from multiple individuals into a large matrix. An efficient sparse learning algorithm is used to find a structural "basis" (dictionary). The dictionary is then decoded to generate representations of actual WM tracts in individuals. We applied our method to the most recent DTI dataset from the Alzheimer's Disease Neuroimaging Initiative (ADNI), including data from 230 participants. With this method, we identified significantly different WM patterns for distinct diagnostic groups. Moreover, the locations of significant WM differences are consistent with prior findings of structural brain abnormalities in Alzheimer's disease (AD).