Detecting Coronary Artery Calcium on Chest Radiographs: Can We Teach an Old Dog New Tricks?
2021; Radiological Society of North America; Volume: 3; Issue: 3 Linguagem: Inglês
10.1148/ryct.2021210123
ISSN2638-6135
Autores Tópico(s)Cardiac Imaging and Diagnostics
ResumoHomeRadiology: Cardiothoracic ImagingVol. 3, No. 3 PreviousNext CommentaryFree AccessCardiovascular Disease PreventionDetecting Coronary Artery Calcium on Chest Radiographs: Can We Teach an Old Dog New Tricks?Sumit Gupta , Ron BlanksteinSumit Gupta , Ron BlanksteinAuthor AffiliationsFrom the Department of Radiology, Cardiovascular Division, Brigham and Women's Hospital, 75 Francis St, Boston, MA 02115.Address correspondence to S.G. (e-mail: [email protected]).Sumit Gupta Ron BlanksteinPublished Online:Jun 17 2021https://doi.org/10.1148/ryct.2021210123MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also article by Kamel et al in this issue.Sumit Gupta, MBBS, PhD, is an attending cardiovascular and thoracic radiologist at Brigham and Women's Hospital and instructor of radiology with the Harvard Medical School. He trained in the United Kingdom and obtained his PhD from University of Leicester. His research interests include the role of cardiac CT in spontaneous coronary artery dissection, myocardial perfusion assessment with cardiac MRI, and deep learning applications in cardiothoracic imaging.Download as PowerPointOpen in Image Viewer Ron Blankstein, MD, MSCCT, is director of cardiac CT and associate director of the Cardiovascular Imaging Program at Brigham and Women's Hospital in Boston, Massachusetts. He is also a professor of medicine at Harvard Medical School. Dr Blankstein is the immediate past president of the Society of Cardiovascular Computed Tomography and serves as an associate editor of JACC: Cardiovascular Imaging and Radiology: Cardiothoracic Imaging.Download as PowerPointOpen in Image Viewer Imaging plays an essential role in cardiovascular medicine, and recent advances in artificial intelligence (AI) are likely to further expand such opportunities (1). AI helps not only in improving the diagnostic performance of existing tests but also identifying imaging findings that may not be obvious, thus improving the detection of various diseases and improving the prognostic value of imaging tests. Machine learning (ML), an application of AI, is based on computer algorithms that automatically learn and improve from experience through analyzing vast amounts of data. Deep learning (DL) is a type of ML that uses computer modeling, known as artificial neural networks, to identify complex relationships in large data sets and make predictions from new input data (2). Deep convolutional neural networks (DCNNs) are the most commonly used DL networks for image analysis.An increasingly used imaging test to enhance cardiovascular disease prevention is non–contrast material–enhanced, electrocardiographically gated CT for coronary artery calcium (CAC) scoring (3,4). CAC scoring has evolved as a highly effective tool to predict all-cause, coronary heart disease, cardiovascular disease, and non–cardiovascular disease mortality (5). Emerging data suggest an expanding role of CAC in younger and low-risk populations (4). Moreover, CAC outperforms all other traditional cardiovascular disease risk factors in predicting long-term all-cause mortality. This has prompted increased recognition regarding the importance of assessing CAC in noncardiac chest CT scans, which are more commonly performed.Chest radiography, being the most commonly performed diagnostic imaging procedure, has the potential to be an even more far-reaching prognostic tool. Yet, it is well known that detecting coronary calcifications on a chest radiograph is a challenging endeavor. In this issue of Radiology: Cardiothoracic Imaging, Kamel et al describe the application of a DL-based algorithm to detect coronary artery calcification on chest radiographs (6).The authors utilized a total of 1689 chest radiographs with paired calcium score CT performed within a 12-month period for training, validation, and testing a DCNN. The authors fine-tuned a pretrained open source network architecture for prediction of (a) zero versus nonzero total calcium scores, (b) presence or absence of calcium among individual coronary artery territories, and (c) discrimination calcium score above or below the thresholds of 100 and 300 on both frontal and lateral chest radiographs.A test set of 241 chest radiographs (approximately 15% of the total data set) was used for binary classification between zero and nonzero Agatston scores, attaining an area under the receiver operating characteristic curve (AUC) of 0.73 on frontal chest radiographs and 0.70 on lateral chest radiographs. The authors reported that the overall performance of detecting CAC decreased when used for assessment of individual coronary arteries. While the discriminatory value of detecting CAC in the left main coronary artery (AUC, 0.64–0.66) and left anterior descending coronary artery (AUC, 0.62) was lower, the performance for detecting calcifications of the right coronary artery on frontal chest radiograph (AUC, 0.71) or the left circumflex artery on lateral chest radiograph (AUC, 0.75) was higher, likely due to better visualization of the respective arteries on corresponding radiographic projections.Findings of this study reinforce the potential of ML methods in detecting findings that may be imperceptible by a reporting radiologist utilizing a simple commonly performed imaging test such as chest radiography.While the majority of current ML research has focused on techniques to estimate the amount of calcium on electrocardiographically gated cardiac CT scans and nongated thoracic CT scans (3), the utilization of chest radiography for this purpose has the advantage of increased opportunities for estimating the risk of having coronary atherosclerosis, as there are billions of chest radiographs (7) obtained annually worldwide. Such “opportunistic” detection of coronary atherosclerosis has the advantage that it can be performed on images that are already acquired for other reasons. However, given the modest diagnostic accuracy of estimating CAC on the basis of chest radiographs, and the fact that future algorithms will likely value sensitivity over specificity, it seems plausible that individuals who are found to have “possible CAC” on their radiograph may then still benefit from undergoing formal CAC testing with CT. This will especially hold true if these people do not have known coronary atherosclerosis, are not on optimal preventive therapies, and are interested in having more information regarding their risk of future atherosclerotic cardiovascular disease (ASCVD) events. The use of downstream CAC testing would thus be based on both the patient's clinical history as well as shared decision-making, in a manner which is analogous to the way in which further evaluation can be considered among women who are found to have breast arterial calcifications (8).One of the major problems with ML algorithms is that the decision-making process is not completely clear, and therefore, they are often perceived as “black boxes.” It is not unreasonable to ask whether an ML algorithm is truly detecting coronary artery calcification on chest radiographs. Recently, ML has been utilized to develop DCNN to predict long-term mortality from findings on a single chest radiograph (9). In a study by Lu et al, attention maps indicating the anatomy contributing to the DCNN included the cardiomediastinal silhouette and chest wall. Similarly, in this study the attention maps were mainly localized to cardiac silhouette; however, other findings on the chest radiograph may have influenced the classification, including noncoronary cardiac calcifications (eg, mitral or aortic valve calcifications) and devices such as cardiac pacemakers or defibrillators. Despite this limitation, the data presented by the authors highlight the immense potential of using DL algorithms on chest radiographs to predict coronary artery calcification. Furthermore, the authors demonstrate that people predicted to have a nonzero Agatston score on frontal radiographs had a significantly higher 10-year ASCVD risk compared with those predicted to have a zero Agatston score (17.2% vs 11.9%). DCNN-assessed nonzero calcification on frontal radiograph was a significant independent predictor of nonzero Agatston score when assessed along traditional ASCVD risk factors in a multivariate logistic regression model.As the authors have discussed, validation of the findings presented in this study is needed, including the use of larger and varied data sets. In addition, there are a few important limitations, as have also been appropriately discussed by the authors. In this study, high resolution Digital Imaging and Communications in Medicine images were converted to lower resolution JPEG images to accommodate the input design of the pretrained DCNN. Future studies with higher resolution images may help in improving the classification of nonzero versus zero total calcium score and differentiate coronary artery calcification from other noncoronary cardiac calcifications. Transparency of current ML algorithms is an issue preventing its acceptance in clinical practice. Although the attention maps generated indicate the anatomy contributing to the DCNN, there is a need for more precise explanation of the individual predictions made by ML algorithms. Such transparent AI systems where the decision-making process is understandable and explainable would lead to greater acceptance by the medical community (10).The potential applications of the current study are substantial, and the authors have proven that in fact you can “teach an old dog new tricks.”Disclosures of Conflicts of Interest: S.G. disclosed no relevant relationships. R.B. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: author is an associate editor for Radiology: Cardiothoracic Imaging (not involved in handling of the article). Other relationships: disclosed no relevant relationships.Authors declared no funding for this work.References1. Lin A, Kolossváry M, Motwani M, et al. Artificial Intelligence in Cardiovascular Imaging for Risk Stratification in Coronary Artery Disease. Radiol Cardiothorac Imaging 2021;3(1):e200512. Link, Google Scholar2. Rajkomar A, Dean J, Kohane I. Machine Learning in Medicine. N Engl J Med 2019;380(14):1347–1358. Crossref, Medline, Google Scholar3. Obisesan OH, Osei AD, Uddin SMI, Dzaye O, Blaha MJ. An Update on Coronary Artery Calcium Interpretation at Chest and Cardiac CT. Radiol Cardiothorac Imaging 2021;3(1):e200484. Link, Google Scholar4. Orringer CE, Blaha MJ, Blankstein R, et al. The National Lipid Association scientific statement on coronary artery calcium scoring to guide preventive strategies for ASCVD risk reduction. J Clin Lipidol 2021;15(1):33–60. Crossref, Medline, Google Scholar5. Adelhoefer S, Uddin SMI, Osei AD, Obisesan OH, Blaha MJ, Dzaye O. Coronary Artery Calcium Scoring: New Insights into Clinical Interpretation-Lessons from the CAC Consortium. Radiol Cardiothorac Imaging 2020;2(6):e200281. Link, Google Scholar6. Kamel PI, Yi PH, Sair HI, Lin CT. Prediction of Coronary Artery Calcium and Cardiovascular Risk on Chest Radiographs Using Deep Learning. Radiol Cardiothorac Imaging 2021;3(3):e200486. Link, Google Scholar7. World Health Organization. Communicating radiation risks in paediatric imaging. http://www.who.int/ionizing_radiation/pub_meet/radiation-risks-paediatric-imaging/en/. Accessed April 27, 2021. Google Scholar8. Osman M, Regner S, Osman K, et al. Association Between Breast Arterial Calcification on Mammography and Coronary Artery Disease: A Systematic Review and Meta-Analysis. J Womens Health (Larchmt) 2021. https://doi.org/10.1089/jwh.2020.8733. Published online April 7, 2021. Google Scholar9. Lu MT, Ivanov A, Mayrhofer T, Hosny A, Aerts HJWL, Hoffmann U. Deep Learning to Assess Long-term Mortality From Chest Radiographs. JAMA Netw Open 2019;2(7):e197416. Crossref, Medline, Google Scholar10. Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. Causability and explainability of artificial intelligence in medicine. Wiley Interdiscip Rev Data Min Knowl Discov 2019;9(4):e1312. Crossref, Medline, Google ScholarArticle HistoryReceived: Apr 28 2021Revision requested: May 18 2021Revision received: Apr 28 2021Accepted: May 18 2021Published online: June 17 2021 FiguresReferencesRelatedDetailsAccompanying This ArticlePrediction of Coronary Artery Calcium and Cardiovascular Risk on Chest Radiographs Using Deep LearningJun 17 2021Radiology: Cardiothoracic ImagingRecommended Articles Prediction of Coronary Artery Calcium and Cardiovascular Risk on Chest Radiographs Using Deep LearningRadiology: Cardiothoracic Imaging2021Volume: 3Issue: 3Automated Coronary Artery Calcium Scoring for Chest CT ScansRadiology2020Volume: 295Issue: 1pp. 80-81Progression of Aortic Valve Calcification and Coronary Atherosclerosis: Similar but Not the SameRadiology2021Volume: 300Issue: 1pp. 87-88Deep Learning–Quantified Calcium Scores for Automatic Cardiovascular Mortality Prediction at Lung Screening Low-Dose CTRadiology: Cardiothoracic Imaging2021Volume: 3Issue: 2Artificial Intelligence in Cardiovascular Imaging for Risk Stratification in Coronary Artery DiseaseRadiology: Cardiothoracic Imaging2021Volume: 3Issue: 1See More RSNA Education Exhibits Coronary Artery Calcium Scoring: Past, Present and FutureDigital Posters2020Cardiac Image Analysis with Deep Learning MethodsDigital Posters2018Spontaneous Coronary Artery Dissection (SCAD): Emergency Radiology PerspectiveDigital Posters2019 RSNA Case Collection Internal carotid artery dissectionRSNA Case Collection2021Aortopulmonary fistula due to aortic dissection RSNA Case Collection2020Anomalous origin of the right coronary artery from the left sinus of ValsalvaRSNA Case Collection2020 Vol. 3, No. 3 Metrics Altmetric Score PDF download
Referência(s)