Fine-grained spatial information extraction in radiology as two-turn question answering.

OBJECTIVES

Radiology reports contain important clinical information that can be used to automatically construct fine-grained labels for applications requiring deep phenotyping. We propose a two-turn question answering (QA) method based on a transformer language model, BERT, for extracting detailed spatial information from radiology reports. We aim to demonstrate the advantage that a multi-turn QA framework provides over sequence-based methods for extracting fine-grained information.

METHODS

Our proposed method identifies spatial and descriptor information by answering queries given a radiology report text. We frame the extraction problem such that all the main radiology entities (e.g., finding, device, anatomy) and the spatial trigger terms (denoting the presence of a spatial relation between finding/device and anatomical location) are identified in the first turn. In the subsequent turn, various other contextual information that acts as important spatial roles with respect to a spatial trigger term are extracted along with identifying the spatial and other descriptor terms qualifying a radiological entity. The queries are constructed using separate templates for the two turns and we employ two query variations in the second turn.

RESULTS

When compared to the best-reported work on this task using a traditional sequence tagging method, the two-turn QA model exceeds its performance on every component. This includes promising improvements of 12, 13, and 12 points in the average F1 scores for identifying the spatial triggers, Figure, and Ground frame elements, respectively.

DISCUSSION

Our experiments suggest that incorporating domain knowledge in the query (a general description about a frame element) helps in obtaining better results for some of the spatial and descriptive frame elements, especially in the case of the clinical pre-trained BERT model. We further highlight that the two-turn QA approach fits well for extracting information for complex schema where the objective is to identify all the frame elements linked to each spatial trigger and finding/device/anatomy entity, thereby enabling the extraction of more comprehensive information in the radiology domain.

CONCLUSION

Extracting fine-grained spatial information from text in the form of answering natural language queries holds potential in achieving better results when compared to more standard sequence labeling-based approaches.