A Hierarchical Approach for Extracting and Displaying Entities and Relations from Radiology Medical Reports

Extracting information from medical reports can be challenging due to the large volume of data. Therefore, this study proposes a method that uses a hierarchical classification approach with two levels, each consisting of a neural network instance. One for extracting clinical anatomical or observational entities along with their levels of uncertainty, and another for classifying the relations that exist between these entities. For this research, 600 radiological reports from the RadGraph dataset were used. The entity extraction task achieved an F1-score of 91%, while the entity classification and relation classification tasks achieved 88% each. Our hierarchical method enhances entity and relation classification performance by filtering and double checking classified entries.