Cerebral imaging is now acknowledged as a crucial research topic in Psychiatry. However, a gap remains between scientific results and clinical applications. For example, a large number of studies have focused on statistical associations with a disease, symptoms or treatment effects on a cross-sectional design. Results are thus informative at a specific time point whereas the disease and its cerebral phenotypes change overtime. Longitudinal imaging enables to identify brain structures and functions changes over time but requires specific preprocessing to avoid bias such as interpolation and registration asymmetries [1]. By creating a midpoint average image, patients’ scans are equally manipulated and statistics are unlikely to be biased.

So far, cerebral imaging do not provide information on diagnosis and/or prognosis and clinicians do not use cerebral imaging in everyday practice. However, recent improvements in modeling cerebral imaging data using multivariate statistics and pattern recognition (i.e. machine learning) might offer the possibility to use imaging in clinical settings. Indeed, it has been shown that machine learning enables to distinguish patients with depressive disorders to controls based on cerebral activation during sad faces visualization [2]. Using a prognostic approach, Tognin et al. [3] were able to predict symptoms progression based on cortical thickness among ultra-high risk for psychosis. However, these applications need to be carefully interpreted in order to preclude inflated optimism [4]. On the basis on this literature, we propose to expose our preliminary results based on combining basal arterial spin labeling and diffusion tensor imaging to improve diagnosis performances of depression in 30 patients and controls.