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

Area of Science:

• Natural Language Processing
• Medical Informatics
• Radiology

Background:

• Radiology reports contain crucial clinical data for deep phenotyping.
• Extracting fine-grained spatial information from these reports is challenging.
• Existing sequence-based methods have limitations in capturing complex relationships.

Purpose of the Study:

• To propose a two-turn question answering (QA) method using BERT for detailed spatial information extraction from radiology reports.
• To demonstrate the superiority of a multi-turn QA framework over sequence-based methods for fine-grained information extraction.
• To enable automated construction of detailed labels for deep phenotyping applications.

Main Methods:

• A transformer-based language model, BERT, is employed for a two-turn QA approach.
• The first turn identifies key radiology entities (finding, device, anatomy) and spatial triggers.
• The second turn extracts contextual information and descriptors related to spatial roles and entities, using query templates.

Main Results:

• The proposed two-turn QA model significantly outperformed the best-reported sequence tagging method on all components.
• Average F1 scores showed improvements of 12, 13, and 12 points for spatial triggers, Figure, and Ground frame elements, respectively.
• Incorporating domain knowledge in queries enhanced results for specific spatial and descriptive elements, particularly with a clinical BERT model.

Conclusions:

• The two-turn QA approach effectively extracts comprehensive spatial and descriptive information from radiology reports.
• This method is well-suited for complex schemas requiring identification of all frame elements linked to entities and spatial triggers.
• Natural language query-based extraction shows significant potential compared to standard sequence labeling approaches in the radiology domain.