Domain specific word embeddings for natural language processing in radiology.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Growing interest in machine learning for radiology NLP.
• Existing models often use general web corpora embeddings due to a lack of radiology-specific data.
• This limits the performance of NLP tasks on radiological text.

Purpose of the Study:

• To evaluate Radiopaedia as a corpus for creating radiology-specific word embeddings.
• To enhance NLP task performance using these specialized embeddings.
• To improve the analysis of radiological text.

Main Methods:

• Trained GloVe word embeddings (50-300D) on Radiopaedia articles.
• Evaluated embeddings on analogy completion tasks.
• Used a shallow neural network with Radiopaedia and Wikipedia embeddings for article labeling.
• Assessed performance using exact match accuracy and Hamming loss with statistical tests.

Main Results:

• Radiopaedia embeddings showed superior performance in specific analogy tasks (tumor origin, organ adjectives).
• Radiopaedia-based models significantly outperformed Wikipedia-based models in article labeling accuracy and Hamming loss across all tested dimensions.
• Statistical significance was confirmed for key performance improvements.

Conclusions:

• Developed and validated radiology-specific word embeddings from Radiopaedia.
• Demonstrated that these embeddings preserve medical semantics and boost NLP task performance.
• Highlighted the benefit of cultivating specialized corpora for future radiology NLP models.