Automatic Normalization of Anatomical Phrases in Radiology Reports Using Unsupervised Learning

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Radiology reports are predominantly free-text, making information retrieval challenging and time-consuming.
• Manual review of prior reports for diagnostic recall is inefficient and prone to human error.
• Automated structuring of radiology reports is needed to streamline clinical workflows.

Purpose of the Study:

• To develop an unsupervised machine learning approach for automatic structuring of radiology reports.
• To detect and normalize anatomical phrases using the Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT) ontology.
• To improve the efficiency and accuracy of accessing critical patient information from radiology reports.

Main Methods:

• Combined word embedding-based semantic learning with ontology-based concept mapping.
• Trained a word embedding model on a large corpus of unlabeled radiology reports.
• Utilized 56 anatomical labels from SNOMED CT for concept normalization.

Main Results:

• The proposed unsupervised approach achieved an average precision of 82.6%.
• Outperformed several state-of-the-art supervised and unsupervised methods in concept normalization.
• Demonstrated superior performance compared to conventional approaches for concept normalization.

Conclusions:

• The unsupervised framework effectively structures radiology reports by normalizing anatomical terms.
• Leverages semantic learning via word embeddings, mitigating the need for large annotated datasets.
• Offers a scalable and efficient solution for enhancing radiology report usability in clinical practice.