Deep Learning and Risk Assessment in Acute Pulmonary Embolism

HomeRadiologyVol. 302, No. 1 PreviousNext Reviews and CommentaryFree AccessEditorialDeep Learning and Risk Assessment in Acute Pulmonary EmbolismAndetta R. Hunsaker Andetta R. Hunsaker Author AffiliationsFrom the Department of Radiology, Brigham and Women’s Hospital, 75 Francis St, Boston, MA 02115.Address correspondence to the author (e-mail: [email protected]).Andetta R. Hunsaker Published Online:Sep 28 2021https://doi.org/10.1148/radiol.2021211897MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Fink et al in this issue.Dr Andetta Hunsaker is an associate professor in radiology at Harvard Medical School and division chief of Thoracic imaging at Brigham and Women’s Hospital in Boston. Her research interests include pulmonary embolism and pulmonary hypertension. She has authored and coauthored several peer-reviewed papers in this area.Download as PowerPointOpen in Image Viewer Acute pulmonary embolism (PE) is a major cause of morbidity and mortality worldwide. With an annual incidence rate of 39–115 per 100 000 people (1), it is important not only to make the diagnosis in a timely manner but to assess the severity of illness. To this end, many clinical and imaging tools converge to facilitate risk and severity assessments. These tools include laboratory, clinical presentation, and imaging. By using these tools, the aim is to quickly triage patients into medical, surgical, or combination therapies. CT pulmonary angiography is one of the most important imaging tools to evaluate and assess risk stratification and illness severity in these patients. Right ventricular (RV) dysfunction is a parameter that many believe to be a predictor of outcomes, namely all-cause and PE-related mortality (1–5). Several methods can be used to measure RV dysfunction. These methods include RV-to–left ventricular (LV) ratio, clot burden, interventricular septal straightening or bowing, contrast agent reflux into the inferior vena cava, and size of the main pulmonary artery to name some. The issue in imaging is which parameters are predictive of patient outcomes, both short and long term, and which are easily reproducible by all who interpret these studies.In this issue of Radiology, the retrospective study by Fink et al (6) data mined highly structured radiology reports with the help natural language processing (NLP) to assess clot burden as a measure of right heart strain. By using this tool, they developed a PE scoring system for CT pulmonary angiography that provides quantification of right heart strain comparable to the Qanadli clot burden scoring system (7). Their study included 304 patients with PE structured reports. The structured reports contained sufficient descriptions of clot location and degree of occlusion to perform the necessary “script-based” calculation of clot burden score and validation. Data extracted from a PE module gave adequate information such that they were able to adapt their scores to the pulmonary artery obstruction index described by Qanadli et al (7). By using a mathematical formula, the occurrence of PE at the main pulmonary artery and lobar levels was calculated and compared with the Qanadli scoring system (7).The results from Fink et al showed that subgrouping the cohort on the basis of the level of PE occurrence revealed a high positive correlation for the obstruction scores attained by their structured reporting model and that of Qanadli at the main pulmonary artery level (r = 0.90) and high positive correlations at lobar and segmental and subsegmental levels (r = 0.83 and 0.70, respectively) (each P < .001). When assessing severity levels, there was no evidence of differences in areas under the receiver operating characteristic curve between the two scoring methods (P = .68) (6). Additionally, they found time savings by using the structured reporting deep learning algorithm with an average savings of 1.3 minutes versus 3.0 minutes (P < .001), respectively, compared with the Qanadli scoring system. Fink et al concluded that they were able to successfully implement text mining to extract key features from structured reports and develop a “simplified CT angiography scoring system for PE severity that strongly correlated with the Qanadli score as a conventional clot burden index (r = 0.94; P < .001).” Additionally, they found higher confidence levels (4.2 vs 3.6; P < .001) and interobserver agreement with their system compared with the Qanadli scoring system.This study has two interesting areas for discussion. One area is the evaluation of CT pulmonary angiography for PE and, most importantly, the role of the radiologist in participating in patient care. As radiologists, an important point to keep in mind is what concerns the clinician at the bedside as they evaluate a patient with an acute PE. The second area is that of what to do with deep learning techniques as they relate to radiology and, in this case, acute PE.The first point of discussion is the evaluation of CT pulmonary angiography images in acute PE. Many parameters have been studied for possible predictors of outcome and severity of acute PE. The goal of any CT pulmonary angiography evaluation should be to ascertain severity of disease in the context of the clinical implications of the findings noted, which then have bearing on outcome. In a study of 635 CT pulmonary angiography studies that were positive for acute PE, Furlan et al (2) explored the correlation between the direct volumetric measurement of clot, the semiquantitative clot burden scoring systems of Qanadli and Mastora, and signs of right heart dysfunction at CT pulmonary angiography to determine whether clot burden and signs of right heart dysfunction were associated with short-term mortality. To achieve this, they looked at several signs of right heart strain at CT pulmonary angiography to ascertain which parameter was most indicative of right heart dysfunction. The signs of right heart strain they evaluated included RV-to-LV ratio, pulmonary artery to aortic ratio, interventricular septal bowing or flattening, pulmonary infarcts, and a few other parameters in patients with normal cardiac and respiratory reserve and those with impaired cardiac and/or respiratory reserve.Interestingly, Furlan et al found that clot volume showed only weak correlation with other CT findings of right heart dysfunction. Additionally, at univariable analysis, they found significant association with short-term mortality with increased RV-to-LV ratio (P < .001), increased pulmonary artery-to-aortic ratio (P = .052), flattening of interventricular septum (P = .03), bowing of interventricular septum (P = .005), and contrast medium reflux into the inferior vena cava (P = .001). No significant association was found between the short-term mortality and clot burden (P = .75). On multivariable analysis, history of congestive heart failure (P = .004), cancer (P = .004), and increased RV-to-LV ratio (P < .001) were the only factors that independently showed an association with short-term mortality. Similar results were found by Meinel et al (5) in a meta-analysis of 49 studies and 13 162 patients. Meinel et al showed that of all the parameters measured, increased RV-to-LV ratio measured on transverse CT images demonstrated the highest risk, with a 2.5-times higher all-cause mortality and five times higher risk for PE-related mortality. Clot burden was not predictive of all-cause mortality.The second point of discussion area is that of deep learning tools. There are several tools for applying deep learning to research and clinical practice in the field of radiology. Some deep learning tools require structured reports, as Fink et al reported in this issue of Radiology, and others require unstructured reports. For example, Chen et al (8) compared the convolutional neural network (CNN) model with a traditional NLP model in CT pulmonary angiography unstructured reports to evaluate their performance in extracting PE findings in the reports. They found that the CNN model had an accuracy that either exceeded or was equivalent to that of the traditional NLP and was more generalizable compared with NLP systems that require structured reports (8). One important difference between NLP and CNN models is that NLP models require specific term definitions, appropriate syntax, and proper coding of terms. Thus, NLP techniques are more laborious when compared with a CNN model, such as that used by Chen et al, which requires no previous definition of terms or specific words relative to descriptive findings in PE. Thus, CNN models are more generalizable and easier to use. Fink et al (6) recognized this limitation of their technique and stated that they were able to use normal expressions as patterns of formation of words and phrases to mine structured reports for PE descriptors that would indicate clot site and degree of pulmonary artery occlusion (6). Though laudable, this is the drawback of the technique used, and detection of unexpected semantic variations is the downside. However, this is a fascinating area for research.The aim of the study by Fink et al (6) was to demonstrate that structured reporting and NLP techniques can be used as a tool for answering scientific questions and not just for improved reporting. To this end, the authors developed a PE scoring system on the basis of CT pulmonary angiography that has major implications, including prediction of risk of morbidity and mortality in patients who present with acute PE. As we answer scientific questions, it may be well to combine these questions with clinical assessment. Patient outcome does not depend simply on the size of the PE. Rather, CT pulmonary angiography along with the patient’s cardiopulmonary status and laboratory markers of myocardial injury (eg, plasma troponin and plasma levels of natriuretic peptides, such as B-type natriuretic peptides) are key in assessing the risk of short- and long-term mortality and the path of treatment. The use of deep learning techniques synchronously in all these areas could be a powerful tool.Disclosures of Conflicts of Interest: A.R.H. disclosed royalties from Wolters Kluwer.References1. Konstantinides SV, Meyer G, Becattini C, et al. 2019 ESC Guidelines for the diagnosis and management of acute pulmonary embolism developed in collaboration with the European Respiratory Society (ERS): The Task Force for the diagnosis and management of acute pulmonary embolism of the European Society of Cardiology (ESC). Eur Respir J 2019;54(3):1901647. Crossref, Medline, Google Scholar2. Furlan A, Aghayev A, Chang CC, et al. Short-term mortality in acute pulmonary embolism: clot burden and signs of right heart dysfunction at CT pulmonary angiography. Radiology 2012;265(1):283–293. Link, Google Scholar3. Subramaniam RM, Mandrekar J, Chang C, et al. Pulmonary embolism outcome: a prospective evaluation of CT pulmonary angiographic clot burden score and ECG score. AJR Am J Roentgenol 2008;190(6):1599–1604. Crossref, Medline, Google Scholar4. Beenen LFM, Bossuyt PMM, Stoker J, Middeldorp S. Prognostic value of cardiovascular parameters in computed tomography pulmonary angiography in patients with acute pulmonary embolism. Eur Respir J 2018;52(1):1702611. Crossref, Medline, Google Scholar5. Meinel FG, Nance JW Jr, Schoepf UJ, et al. Predictive value of computed tomography in acute pulmonary embolism: systematic review and meta-analysis. Am J Med 2015;128(7):747–59.e2. Crossref, Medline, Google Scholar6. Fink M, Mayer VL, Schneider T, et al. CT angiography clot burden score from data mining of structured reports for pulmonary embolism. Radiology 2021.https://doi.org/10.1148/radiol.2021211013. Published online September 28, 2021. Link, Google Scholar7. Qanadli SD, El Hajjam M, Vieillard-Baron A, et al. New CT index to quantify arterial obstruction in pulmonary embolism: comparison with angiographic index and echocardiography. AJR Am J Roentgenol 2001;176(6):1415–1420. Crossref, Medline, Google Scholar8. Chen MC, Ball RL, Yang L, et al. Deep learning to classify radiology free-text reports. Radiology 2018;286(3):845–852. Link, Google ScholarArticle HistoryReceived: July 26 2021Revision requested: Aug 10 2021Revision received: Aug 11 2021Accepted: Aug 16 2021Published online: Sept 28 2021Published in print: Jan 2022 FiguresReferencesRelatedDetailsAccompanying This ArticleCT Angiography Clot Burden Score from Data Mining of Structured Reports for Pulmonary EmbolismSep 28 2021RadiologyRecommended Articles Pulmonary Embolism at CT Pulmonary Angiography in Patients with COVID-19Radiology: Cardiothoracic Imaging2020Volume: 2Issue: 4Pulmonary Embolism in Hospitalized Patients with COVID-19: A Multicenter StudyRadiology2021Volume: 301Issue: 3pp. E426-E433Stratification, Imaging, and Management of Acute Massive and Submassive Pulmonary EmbolismRadiology2017Volume: 284Issue: 1pp. 5-24CT Angiography Clot Burden Score from Data Mining of Structured Reports for Pulmonary EmbolismRadiology2021Volume: 302Issue: 1pp. 175-184Contemporary Management of Acute Pulmonary Embolism: Evolution of Catheter-based TherapyRadioGraphics2022Volume: 42Issue: 6pp. 1861-1880See More RSNA Education Exhibits Acute Pulmonary Embolism: The Role of the Radiologist from Diagnosis to TreatmentDigital Posters2018Acquired and Congenital Pulmonary Artery Pathologies: Thinking Beyond the Embolism  Digital Posters2018Collateral Damage: CTA of Acute Pulmonary Embolism and Its Correlation with Evolution of COVID-19 PneumoniaDigital Posters2020 RSNA Case Collection Pulmonary Embolism with Right Heart StrainRSNA Case Collection2020Pulmonary embolism with right heart strain RSNA Case Collection2020Pulmonary embolism with right ventricle strainRSNA Case Collection2020 Vol. 302, No. 1 Metrics Altmetric Score PDF download