Brian Avants1, Phil Cook1, Lyle
Ungar1, James Gee1,
We present a novel, unsupervised method, sparse canonical correlation analysis for neuroimaging (SCCAN), that automatically locates correlated sets of voxels in complementary imaging modalities. The method reveals significant and syndrome-specific cortical thickness-diffusion tensor imaging networks in two neurodegenerative diseases, AD and FTD. Subject diagnosis was confirmed by autopsy or CSF-biomarker ratios. The SCCAN summary correlates, in AD, with MMSE reduction and, in FTD, with reduced verbal fluency. Thus, SCCAN identifies disease-specific networks of effects in white matter and cortical thickness that appear in anatomy suspected to be involved in these diseases and that relate specifically to impaired cognitive processes.