Machines capable of analysing and interpreting medical scans with super-human performance are within reach. Deep learning, in particular, has emerged as a promising tool in our work on automatically detecting brain damage. But getting from the lab into clinical practice comes with great challenges. How do we know when the machine gets it wrong? Can we predict failure, and can we make the machine robust to changes in the clinical data? We will discuss some of our most recent work that aims to address these critical issues and demonstrate our latest results on deep learning for analysing medical scans.