Studies have revealed volumetric and connectivity abnormalities in brains of children with autism spectrum disorders (ASD). However, due to the complexity of ASD findings from these studies have been inconsistent. A possible way to improve the robustness of neuroimaging quantification is to introduce biological regularization. Motivated by previous studies on gray matter (GM) diffusivity, which can be used as surrogate biomarker for cortical volume deficits, we proposed a new network-regularized GM diffusivity analysis method to detect the GM diffusivity in the context of white matter (WM) connectivity. Our rationale is that GM diffusivity of anatomically connected regions has similar contributions to group differences. Experiments on simulated and clinical datasets showed better performance of the proposed method both on identification of atypical GM regions and classification of ASD and controls, compared with other methods without network regularization.