Narrative review of open source, proprietary, and experimental artificial intelligence algorithms in radiology

Abstract

Background and Objective: The aim of this study is to educate the reader on the applications of artificial intelligence (AI)-based algorithms, their basic functioning mechanism, and the efficacy of various algorithms which may be encountered in clinical practice. From image reconstruction and interpretation to image segmentation and clinical decision making, AI has demonstrated wide applicability in the field of radiology. The following systematic review yielded a comprehensive, but not exhaustive, summary of AI algorithms pertaining to diagnostic radiology. Methods: Five databases, including MEDLINE, Google Scholar, Association for Computing Machinery (ACM) Digital Library, Institute of Electrical and Electronics Engineers (IEEE) Xplore, and PubMed, were queried for recent research articles pertaining to various aspects of AI in Radiology. Specific criteria narrowed the yield of research articles as those pertaining to AI as applied to image reconstruction, image interpretation, and clinical decision making. Key Content and Findings: A broad overview of AI-driven algorithms encompassing all aspects of radiology from image reconstruction to image interpretation to decisions on next steps are included in this review article. Conclusions: Although AI demonstrates excellent efficiency and efficacy when applied to image reconstruction, current technological limitations hinder the widespread adoption of AI in image interpretation and clinical decision making. Since many algorithms have had recurrent false positive results, integration of AI into the radiologists’ workflow at this precise moment in time does not improve efficiency and accuracy when compared to traditional approaches which do not rely on computer vision and image feature extraction. As AI drives further advancement in the field of radiomics, AI system will become more accurate. In the meantime, false-positive results can be alleviated by confirming the algorithm decision with expert opinion based off the clinical history and physical exam findings given. In that sense, it could act as a safety net for potentially overlooked diagnoses, but not as the final arbiter of diagnosis itself.