Context-Aware Filtering of Unstructured Radiology Reports by Anatomical Region.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Radiology reports contain crucial clinical data but are often unstructured free text.
• Multiple imaging studies in a single report complicate analysis.
• Accurate anatomical classification is vital for downstream AI tasks.

Purpose of the Study:

• To develop a method for sentence-level anatomical classification in radiology reports.
• To improve the creation of structured data for AI models.
• To create a practical and reproducible approach for diverse clinical settings.

Main Methods:

• Developed context-aware classical models with feature engineering.
• Leveraged the precision of clinical language despite formatting variations.
• Validated model generalizability on MIMIC-IV radiology reports.

Main Results:

• Context-aware classical models outperformed neural networks and pre-trained language models.
• Achieved near-optimal performance with minimal labeled training data.
• Demonstrated effective generalization of learned model weights.

Conclusions:

• Context-aware classical models offer a practical and reproducible solution for anatomical classification in radiology reports.
• The approach enhances data quality for AI applications in radiology.
• Effective for downstream tasks like structured label extraction and model training data generation.