Portal de Periódicos da CAPES

Improving Our Understanding of Indolent Lesions: A New Role for AI

2022; Radiological Society of North America; Volume: 4; Issue: 2 Linguagem: Inglês

10.1148/ryai.210312

2638-6100

Steven C. Horii,

Lung Cancer Diagnosis and Treatment

HomeRadiology: Artificial IntelligenceVol. 4, No. 2 PreviousNext CommentaryFree AccessImproving Our Understanding of Indolent Lesions: A New Role for AISteven C. Horii Steven C. Horii Author AffiliationsFrom the Department of Radiology, University of Pennsylvania Perelman School of Medicine, 3400 Spruce St, Philadelphia, PA 19104-4385.Address correspondence to the author (e-mail: [email protected]).Steven C. Horii Published Online:Feb 2 2022https://doi.org/10.1148/ryai.210312MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also article by Yamashita et al in this issue.Steven C. Horii, MD, FACR, FSIIM, received a bachelor of arts degree in natural sciences from Johns Hopkins University in 1972 and his MD from NYU Langone School of Medicine in 1976. After completing a residency in radiology and an abdominal imaging fellowship at NYU Medical Center, he served as assistant professor, and later, associate professor at NYU Medical Center (1980–1988). He served as associate professor of radiology and clinical director of the Image Management and Communications division at Georgetown University Hospital (1988–1992) and as professor (1992–2021) and professor emeritus (2021–present) at the University of Pennsylvania. He is an elected fellow of the American College of Radiology, Society for Imaging Informatics in Medicine, International Society for Optics and Photonics, and College of Physicians of Philadelphia. In 2020, he received the Gold Medal of the Society for Imaging Informatics in Medicine.Download as PowerPointOpen in Image Viewer Yamashita and coauthors have applied natural language processing (NLP) methods to identify abdominal CT and MRI reports that indicated a pancreatic cystic lesion (PCL) was present and then automatically extract the measurements of the identified lesions (1). The results were compared against expert radiologists who reviewed a set of 1000 reports to annotate them regarding the presence or absence of a PCL and the measurements of the cystic lesion(s), if present. In further analysis, follow-up studies among patients with PCLs were evaluated for any change in size of the lesions. The histologic findings of resected lesions were correlated with the size of the lesions and any increase or decrease in size that occurred between the initial and follow-up scans.As the authors note, PCLs present a clinical dilemma. Because PCLs are found as incidental findings in as many as 19.5% of abdominal MRI studies, the question remains as to how these lesions should be managed (2). The American College of Radiology (ACR) Incidental Findings Committee developed an algorithm for clinical management (3). The authors’ development of an NLP system to identify radiology reports (and hence, the studies that generated them) and extract measurements provides for a potential way to accumulate a large database of PCLs and then mine that database to improve understanding of the natural history of PCLs.One finding of the authors’ study took advantage of reports that had been generated prior to the ACR white paper on the management of incidental pancreatic cysts and compared the measurements with reports generated after the publication of the white paper. Their algorithms achieved higher performance in terms of agreement with the radiologists on reports generated after the white paper publication (table 3 of the authors’ article), likely because the radiologists began using the recommended measurement methods described in the white paper and so were more uniform in the reports than when measurements were ad hoc.Overall, the NLP methods the authors used showed a very high degree of agreement with the radiologists who were used as the reference standard for the identification and measurements of PCLs in the reports. The authors also have used publicly available software and made available their modifications and models using the software. Part of their objective was to provide other researchers with the methods used in their article to encourage expansion of their study or application to other NLP tasks.Although the article is primarily a successful demonstration of the use of NLP for a clinically challenging task, the potential for the use of NLP in radiology is extensive, and some aspects are likely already in wide use. A review of NLP in radiology by Pons et al provides both an overview of NLP methods and a list of NLP applications in use at the time of the publication (4). Examples include a system that extracts information from reports and automatically generates Breast Imaging Reporting and Data System (ie, BI-RADS) final assessment categories; a system for automatic extraction of information from clinical free text in electronic medical records; a commercial automated system for coding radiology reports for billing and revenue management; and a system that extracts, structures, and classifies unstructured radiology reports, among many others.The latter example, classifying unstructured radiology reports, has potential economic impact for radiology. The ideas behind structured reporting have been discussed among academic and nonacademic radiologists for years. While benefits have been touted, there are also studies that have shown that structured reporting, in some instances, decreased radiology report accuracy and completeness (5). The Johnson et al article, however, has weaknesses because the reporting system used was then new (in beta testing at the time) and the residents had limited training time using it. In a more recent review, Ganeshan and colleagues provide a comprehensive review of the importance of structured reporting to reduce missed findings, increase confidence of surgeons in resectability of mass lesions, and decrease potential revenue loss due to failure to mention anatomy that should be included in the report (6). Advances in high-speed computing elements have led to real-time or near real-time NLP. Examples are readily found in the commercial sector, such as the popular Zillow application for renting or purchasing real estate (7). The possibility of NLP transforming free-text or partially structured text (reports with standard section headers) in near real time has intriguing implications for report generation, which may eliminate radiologists’ complaints about structured reporting being more time-consuming or constraining than conventional dictation and may reduce the speech recognition errors that plague current systems.Being able to extract information from radiology reports and other electronic medical records is also increasingly meaningful for the quality metrics that are already part of the Merit-based Incentive Payment System (MIPS), which is part of the Medicare Access and Children's Health Insurance Program Reauthorization Act. As Ganeshan et al describe, proposed MIPS quality measures will be required in radiology reports. The Center for Medicare and Medicaid Services has published the “Radiology Preferred Specialty Measure Set,” which describes how the elements are to be reported (through claims or a registry), as well as the measures themselves and the “National Quality Strategy Domain” to which they belong (8). The importance to radiology (and for other health care specialties as well) is that reporting the required elements will be necessary to be reimbursed. Having to extract these required measures manually would add to the cost of complying, and failure to comply (ie, to meet performance thresholds) means a “negative payment adjustment.”A less potentially punitive view of NLP involves the rapid growth of bioinformatics. As we expand the role of radiomics, the potential for NLP coupled with machine learning and other artificial intelligence applications in research can help correlate information across the increasingly complex realm of the other “-omics” fields—genomics, proteomics, and the like. The amount of information is so large that trying to gather diagnostically, therapeutically, and prognostically important data is nearly impossible without assistance from computer-based systems. Even if it were possible to gather information across multiple complex systems, a difficulty arises from the “we don't know what we don't know” problem. Without a way to find, select, correlate, and analyze the information, it would be like having a library with randomly shelved books: The information is there, but how do you find it? More importantly, how do you correlate that information and potentially discover questions and answers about which you had no inkling?Yamashita and colleagues have provided a carefully done study of the use of NLP for a clinically relevant problem. It should serve as an example of what can be done using NLP and as an introduction to the potential for other current and future applications in radiology.Disclosures of Conflicts of Interest: S.C.H. Royalties from the International Society for Optics and Photonics (SPIE) for serving as an editor for the Handbook of Medical Imaging, vol 3, “Display and PACS”, SPIE Press, 2000, no relation to the authors of the article this commentary is in response to, nor any relationship with the Radiological Society of North America (RSNA) on publications; expert witness for defendant in a patent infringement suit, company involved has no known relationship to the authors of the article this commentary is in response to; chairman of the RSNA Research and Education Foundation Grant Program Committee, non-paid position; served on the RSNA RadioGraphics Informatics Exhibits Panel through the end of 2021, non-paid position; stockholder in Apple Computer Corporation.Author declared no funding for this work.References1. Yamashita R, Bird K, Cheung PYC, et al. Automated identification and measurement extraction of pancreatic cystic lesions from free-text radiology reports using natural language processing. Radiol Artif Intell 2022;4(2):e210092. Link, Google Scholar2. D'Ippolito G. Incidental pancreatic cyst: still a lot of road to cover. Radiol Bras 2018;51(4):V–VII. Crossref, Medline, Google Scholar3. Megibow AJ, Baker ME, Morgan DE, et al. Management of Incidental Pancreatic Cysts: A White Paper of the ACR Incidental Findings Committee. J Am Coll Radiol 2017;14(7):911–923. Crossref, Medline, Google Scholar4. Pons E, Braun LMM, Hunink MG, Kors JA. Natural Language Processing in Radiology: A Systematic Review. Radiology 2016;279(2):329–343. Link, Google Scholar5. Johnson AJ, Chen MY, Swan JS, Applegate KE, Littenberg B. Cohort study of structured reporting compared with conventional dictation. Radiology 2009;253(1):74–80. Link, Google Scholar6. Ganeshan D, Duong PT, Probyn L, et al. Structured Reporting in Radiology. Acad Radiol 2018;25(1):66–73. Crossref, Medline, Google Scholar7. DeLuca A, Karmakar A. Near Real-Time Natural Language (NLP) for Customer Interactions. YouTube. https://www.youtube.com/watch?v=w-qGSyzDL6g. Accessed December 15, 2021. Google Scholar8. Center for Medicare and Medicaid Services. Radiology Preferred Specialty Measure Set. https://www.cms.gov/Medicare/Quality-Initiatives-Patient-Assessment-Instruments/PQRS/Downloads/Radiology_Specialty_Measure_Set.pdf. Accessed December 15, 2021. Google ScholarArticle HistoryReceived: Dec 16 2021Revision requested: Dec 20 2021Revision received: Dec 27 2021Accepted: Jan 5 2022Published online: Feb 02 2022 FiguresReferencesRelatedDetailsAccompanying This ArticleAutomated Identification and Measurement Extraction of Pancreatic Cystic Lesions from Free-Text Radiology Reports Using Natural Language ProcessingDec 22 2021Radiology: Artificial IntelligenceRecommended Articles Morton A. Bosniak, MDRadiology2017Volume: 282Issue: 2pp. 616Multimedia-enhanced Radiology Reports: Concept, Components, and ChallengesRadioGraphics2018Volume: 38Issue: 2pp. 462-482Microwave Ablation versus Surgery for Papillary Thyroid Carcinoma: More Therapeutic Options, More ControversiesRadiology2022Volume: 304Issue: 3pp. 714-715BI-RADS Category 3 Is a Safe and Effective Alternative to Biopsy or Surgical ExcisionRadiology2020Volume: 296Issue: 1pp. 42-43Correlate: A PACS- and EHR-integrated Tool Leveraging Natural Language Processing to Provide Automated Clinical Follow-upRadioGraphics2017Volume: 37Issue: 5pp. 1451-1460See More RSNA Education Exhibits Quality Assessment of the BI-RADS 3 Classification and Utility of a BI-RADS 3 AuditDigital Posters2019Multimodality Imaging Findings of Low-Grade DCIS: Should Biopsy Be Performed in Low-Grade DCIS?Digital Posters2020Conventional and Advanced MRI Evaluation of Hypophysis: Is it Possible to Minimize the Use of Gadolinium with DWI and t2-Mapping Techniques?  Digital Posters2019 RSNA Case Collection Central Pontine MyelinolisisRSNA Case Collection2021Metronidazole induced encephalopathyRSNA Case Collection2022EchinococcosisRSNA Case Collection2020 Vol. 4, No. 2 Metrics Downloaded 113 times Altmetric Score PDF download