Automatic Classification and Entity Relation Detection in Hungarian Spinal MRI Reports

Abstract

A great number of radiologic reports are created each year which incorporate the expertise of radiologists. This knowledge could be exploited via machine understanding. This could provide valuable statistics and visualization of the reports, and as training data, and it could also contribute to later automatic reporting applications. In our current work, we present our first steps toward the machine understanding of clinical reports of the spinal region, written in the Hungarian language. Our system provides an automatic classification and connection detection for various entities in the text. Our classification is achieved via bi-directional long short-term memory and conditional random fields producing 0.87–0.95 F1-score values, while the extraction of connection relies on linguistic analysis and predefined rules. The extracted information is displayed in an easily comprehensible, well-formed tree-structure.