Purpose: Deep learning is increasingly being adapted for natural language processing (NLP), enabling conversion of free text into structured data that can be analyzed. This study aims to review the literature regarding applications for deep learning NLP in radiology.
Materials and Methods: A literature search using MEDLINE, Scopus and Google Scholar was conducted, reviewing all published studies up to May 2019 on deep learning technologies for NLP applied in radiology. The following keywords were used: “radiology”, “natural language processing”, “NLP”, “free text”, “deep learning”, “neural network”, “convolutional neural network”, “CNN”, “recurrent neural network” and “RNN”.
Results: We identified eight relevant publications; all were published between 2017 and 2019. Major deep learning models applied for NLP in radiology are convolutional neural networks (CNN), recurrent neural networks (RNN), and two subtypes of RNN: Long Short-Term Memory (LSTM) and Attention networks.
Current applications for deep learning NLP in radiology include flagging and labeling of diagnoses such PE and fractures, labeling follow-up recommendations, and automatic selection of imaging protocols. Studies that compared traditional NLP and deep learning NLP reported deep learning performance to be equivalent to or better than traditional techniques.
Conclusion: Research and use of deep learning NLP in radiology is expected to increase in coming years. Acquaintance with these concepts and potential applications is important in order to be better prepared for the coming technological changes in our field.
