Uma abordagem hierárquica para extração e exibição de entidades e relações em relatórios radiológicos

Abstract

The extraction of information from medical reports is challenging due to the complexity and volume of data. This study proposes a hierarchical two-level approach using independent neural networks: the first level extracts anatomical and observational entities along with their uncertainty levels, while the second classifies the relations between these entities. Experiments were conducted on 600 reports from the RadGraph dataset, with 500 used for training, testing and validation; and 100 reserved exclusively for testing. On the 50-report MIMIC-CXR test set, the models achieved micro F1-scores of 0.756 for entity recognition and 0.532 for relation extraction. On the 50-report CheXpert test set, not seen during training, the results were 0.742 and 0.461, respectively. These findings demonstrate that the hierarchical approach improves accuracy in both entity recognition and relation classification by filtering inputs through the first model and performing double verification with the second.