Training medical image analysis models requires large amounts of expertly\nannotated data which is time-consuming and expensive to obtain. Images are\noften accompanied by free-text radiology reports which are a rich source of\ninformation. In this paper, we tackle the automated extraction of structured\nlabels from head CT reports for imaging of suspected stroke patients, using\ndeep learning. Firstly, we propose a set of 31 labels which correspond to\nradiographic findings (e.g. hyperdensity) and clinical impressions (e.g.\nhaemorrhage) related to neurological abnormalities. Secondly, inspired by\nprevious work, we extend existing state-of-the-art neural network models with a\nlabel-dependent attention mechanism. Using this mechanism and simple synthetic\ndata augmentation, we are able to robustly extract many labels with a single\nmodel, classified according to the radiologist's reporting (positive,\nuncertain, negative). This approach can be used in further research to\neffectively extract many labels from medical text.\n