Introduction: Vast amounts of medical information about patients’ conditions remain stored in hospitals database systems in the form of textual reports, yet they are often lacking semanticstructure. The main goal of this research was to inspect the feasibility of developing an automated method of labelling radiology reports, as a precursor for building query-capable report databases. Methods: The analysed dataset was composed of 1295 radiology reports concerning the condition of a knee, retrospectively gathered at Clinical Hospital Centre Rijeka, Croatia. Reports were manually labelled with one or more labels from a set of 14 most commonly occurring clinical conditions. Bag-of-words TF-IDF features were extracted after primary preprocessing. 10-fold cross-validation was used to evaluate the performance of several state-of-the-art supervised learning models on the dataset. Accuracy, precision, recall and F1 scores were estimated for each of the models. Error analysis was performed in order to determine most frequently occurring mistakes. Results: Best performance was recorded for a neural network model, having an F1 score of 81%. Support vector machine model also performed very well with an F1 score of 80%. Random forests model and naive Bayes model fared significantly worse with F1 scores of 64% and 62%, respectively. Performance of the neural network and the support vector machine models on more frequently occurring labels was especially promising with F1 scores of 91% and 87% for labels arthrosis and injury, respectively. Conclusion: Supervised machine-learning models are more than capable of dealing with automated labelling of unstructured radiology reports, if a sufficiently large labelled dataset is available.