Linna Nathaniel, Kahn Charles E
Department of Radiology, University of Pennsylvania, 3400 Spruce Street, Philadelphia, PA 19104, USA.
Department of Radiology, University of Pennsylvania, 3400 Spruce Street, Philadelphia, PA 19104, USA; Institute for Biomedical Informatics, University of Pennsylvania, USA.
Int J Med Inform. 2022 Jul;163:104779. doi: 10.1016/j.ijmedinf.2022.104779. Epub 2022 Apr 26.
Recent advances in performance of natural language processing (NLP) techniques have spurred wider use and more sophisticated applications of NLP in radiology. This study systematically reviews the trends and applications of NLP in radiology within the last five years.
A search of three databases of peer-reviewed journal articles and conference papers from January 1, 2016 to April 21, 2021 resulted in a total of 228 publications included in the review. Manuscripts were analyzed by several factors, including clinical application, study setting, NLP technique, and performance.
Of the 228 included publications, there was an overall increase in number of studies published with an increase in use of machine learning models. NLP models showed high performance: >50% of publications reported F1 > 0.91. There was variable sample size across the studies with a median of 3708 data points, most commonly radiology reports. 145 studies utilized data from a single academic center. Applications were classified as clinical (n = 87), technical (n = 66), quality improvement (n = 61), research (n = 9), and education (n = 5).
There has been a continued increase in number of studies involving NLP in radiology. Newer NLP techniques, including word embedding, deep learning, and transformers, are being applied and show improved performance. There has been growth in the interpretative and non-interpretative use of NLP techniques in radiology and has great capacity to improve patient care and delivery. Although the performance and breadth of NLP applications is impressive, there is an overall lack of high-level evidence for actual clinical application of published tools.
NLP applications in radiology has been increasing studied and more accurate in the last 5 years. More direct clinical application and portability of the NLP pipelines is need to reach the technology's full potential.
自然语言处理（NLP）技术性能的最新进展推动了NLP在放射学领域的更广泛应用和更复杂的应用。本研究系统回顾了过去五年中NLP在放射学领域的发展趋势和应用情况。
检索2016年1月1日至2021年4月21日期间三个同行评审期刊文章和会议论文数据库，共筛选出228篇纳入本综述的出版物。对稿件进行了多方面分析，包括临床应用、研究环境、NLP技术和性能。
在纳入的228篇出版物中，随着机器学习模型使用的增加，发表的研究数量总体呈上升趋势。NLP模型表现出高性能：超过50%的出版物报告F1>0.91。各项研究的样本量各不相同，中位数为3708个数据点，最常见的是放射学报告。145项研究使用了来自单一学术中心的数据。应用分为临床（n=87）、技术（n=66）、质量改进（n=61）、研究（n=9）和教育（n=5）。
涉及放射学中NLP的研究数量持续增加。包括词嵌入、深度学习和变换器在内的更新的NLP技术正在得到应用，并显示出更好的性能。NLP技术在放射学中的解释性和非解释性应用都有所增长，并且在改善患者护理和服务方面具有很大潜力。尽管NLP应用的性能和广度令人印象深刻，但对于已发表工具的实际临床应用，总体上缺乏高级别证据。
在过去五年中，NLP在放射学中的应用研究不断增加且更加准确。需要更多直接的临床应用以及NLP管道的可移植性，以充分发挥该技术的潜力。
