Related Experiment Video
Updated: Dec 17, 2025
07:53
Author Spotlight: Advancing 3D Modeling for Enhanced Diagnosis and Treatment of Pulmonary Nodules in Early-Stage Lung Cancer
Published on: October 13, 2023
1.9K
Understanding spatial language in radiology: Representation framework, annotation, and spatial relation extraction
Surabhi Datta1, Yuqi Si1, Laritza Rodriguez2
1School of Biomedical Informatics, The University of Texas Health Science Center, Houston, TX, United States.
Journal of Biomedical Informatics
|June 21, 2020
Summary
This study introduces Rad-SpRL, a framework for extracting spatial information from radiology reports. Deep learning models achieved high accuracy in identifying anatomical locations and associated diagnoses, improving clinical informatics.
Area of Science:
- Natural Language Processing (NLP)
- Medical Informatics
- Radiology
Background:
- Radiology reports describe spatial relationships crucial for diagnosis.
- Extracting this spatial information is challenging but vital for clinical applications.
Purpose of the Study:
- To develop a method for extracting spatial representations from radiology reports.
- To define a framework (Rad-SpRL) for encoding spatial information.
Main Methods:
- Annotated 2,000 chest X-ray reports using the Rad-SpRL framework.
- Employed deep learning NLP models (Bi-LSTM-CRF, BERT, XLNet) for information extraction.
- Utilized word and character-level encodings for spatial indicator and role identification.
Main Results:
- Achieved high F1 scores for Spatial Indicator extraction (up to 91.29% with XLNet).
- Obtained strong overall F1 measures for spatial role extraction (92.9% with gold indicators, 85.6% with predicted indicators using BERT).
Conclusions:
- The proposed Rad-SpRL framework and deep learning methods effectively extract clinically significant spatial information from radiology reports.
- This work facilitates downstream clinical informatics applications by structuring unstructured text data.
