Fig. 2.

Deep learning MRI neuroradiology report classifier for automated dataset labelling. Free-text reports are converted into a list of integer word identifiers which are passed into a transformer-based language encoder network. This network converts each word into a 768-dimensional contextualised embedding vector and contains ~110 million parameters which are initialised with weights from BioBERT—a biomedical language model pre-trained on all of English Wikipedia (2.5 billion words), PubMed abstracts (4.5 billion words), and PMC full-text articles (13.5 billion words). A custom attention network aggregates these vectors into a report representation by a taking weighted sum of embedding vectors, with the weight of each word determined by its importance to the classification decision. A fully connected neural network with a single hidden layer then converts this to a probability that the report describes a category of interest, e.g. abnormal, mass, acute stroke. The entire network—i.e. the transformer language model, attention module, and classification network—is trained end-to-end on the basis of the binary cross-entropy between the model predictions and reference-standard report labels using the Adam optimizer