Psychiatric disorders include a broad panel of heterogeneous conditions. Among the most severe psychiatric diseases, in intensity and incidence, depression will affect 15–20% of the population in their lifetime, schizophrenia 0.7–1%, and bipolar disorder 1–2.5%. Today, the diagnosis is solely based on clinical evaluation, causing major issues since it is subjective and as different diseases can present similar symptoms. These limitations in diagnosis lead to limitations in the classification of psychiatric diseases and treatments. There is therefore a great need for new biomarkers, usable at an individual level. Among them, magnetic resonance imaging (MRI) allows to measure potential brain abnormalities in patients with psychiatric disorders. This creates datasets with high dimensionality and very subtle variations between healthy subjects and patients, making machine and statistical learning ideal tools to extract biomarkers from these data. Machine learning brings different tools that could be useful to tackle these issues. On the one hand, supervised learning can support automated classification between different psychiatric conditions. On the other hand, unsupervised learning could allow the identification of new homogeneous subgroups of patients, refining our understanding of the classification of these disorders. In this chapter, we will review current research applying machine learning tools to brain imaging in psychiatry, and we will discuss its interest, limitations, and future applications.