The Potential of Artificial Intelligence to Analyze Chest Radiographs for Signs of COVID-19 Pneumonia

HomeRadiologyVol. 299, No. 1 PreviousNext Reviews and CommentaryFree AccessEditorialThe Potential of Artificial Intelligence to Analyze Chest Radiographs for Signs COVID-19 PneumoniaBram van Ginneken Bram Author AffiliationsFrom the Department Radiology, Radboud University Medical Center, Geert Grooteplein Zuid 10, 6525 GA Nijmegen, Netherlands.Address correspondence author (e-mail: [email protected]).Bram Published Online:Nov 24 2020https://doi.org/10.1148/radiol.2020204238MoreSectionsPDF ToolsImage ViewerAdd favoritesCiteTrack Citations ShareShare onFacebookTwitterLinked In See also article by Wehbe et al in this issue.Bram is professor medical image analysis at Center. He works Fraunhofer MEVIS Bremen, Germany, a founder Thirona, company that develops software provides services analysis. studied physics Eindhoven Technology Utrecht University, where he obtained his doctorate 2001 on computer-aided diagnosis chest radiography. pioneered concept challenges analysis.Download as PowerPointOpen Image Viewer It was about year ago new coronavirus started spread from Wuhan, China. The resulting pandemic unprecedented many ways, one them number scientific publications it has generated. PubMed already lists over 70 000 papers disease 2019 (COVID-19). first publication describing CT appearance pneumonia findings, dates February 6, 2020. To date, Radiology published 40 original research articles topic. These studies have had substantial impact: 12 most cited 2020 (using counts Google Scholar) are all COVID-19, even ranked 12th twice citations 2019. (Interestingly, most-cited applications artificial intelligence [AI].)Of so far, 23 focused CT, only six This likely related fact that, noted Fleischner Society (1) their consensus statement role imaging patient care during pandemic, “chest radiography insensitive mild or early COVID-19.” conclusion based evidence radiographic findings patients with (2). However, countries encourage individuals symptoms consistent quarantine home. such scenario, presenting hospital may more advanced disease, often abnormalities visible radiograph. Because its broad availability, low cost, portability, widely used tool obtain an initial while waiting results molecular diagnostic tests. Radiographic can help assess progression detect other diseases.Many health providers overburdened struggle lack resources interpretation. AI could provide support reading process.In issue (3) present algorithm, coined DeepCOVID-XR, detects single frontal radiographs. not attempt do this. May 2020, Murphy (4) reported validation study CAD4COVID–x-ray, freely available CE-marked commercial solution; I coauthor study. September Zhang (5) introduced CV19-Net. All three algorithms address same task.A strength DeepCOVID-XR trained large multicenter data set: Almost 15 images 4000 cases were positive originating than 20 sites across Northwestern Memorial Health Care System, organization operating Chicago, Ill, region. Nearly anteroposterior bedside system evaluated community hospital. proper external No been train system. A set 300 random (134 COVID-19) presented five radiologists allow direct comparison between human experts.A similar set-up two studies. CAD4COVID–x-ray 454 (223 compared scores radiologists. 416 suspects hospital, although deep learning network pretrained sources. CV19-Net Henry Ford total approximately 5000 (about half which training. negative diagnosed drawback test came hospitals provided training data. On 500 randomly selected images, equally balanced cases, readings radiologists.All systems yielded promising terms area under receiver characteristic curve (AUC), commonly metric continuous output binary classification task. AUC equivalent chance receives higher score set. reached 0.88, comparable (AUC = 0.85, using six-point scale). 0.94, outperforming each readers who did perform scoring. achieved 0.81 slightly outperformed high-sensitivity cutoffs but performed inferior four high specificity cutoff.An interesting strategy pursued design ensemble neural networks diverse characteristics. They different architectures popular today (DenseNet-121, ResNet-50, Inception, Inception-ResNet, Xception, EfficientNet-B2) resolution levels (224 × 224 331 331) fields view (the entire radiograph cropped around automatically segmented lung fields). Research shown zooming lead better performance abnormality detection radiographs (6). study, differences minor, combined approach be robust when applied unseen data.An important avenue further mentioned combine additional input, demographics, vital signs, laboratory simple scoring suggested multimodal increase substantially alone (7). Additionally, predict outcomes guide interventions. predicting intubation mortality just Radiology: (8). Such requires availability sets standardized treatment parameters available. Collecting remains extremely challenging strategies continuously adapted we learn COVID-19.To move forward approaches automated potential, should compare various “nets” now published. Sharing would facilitate making publicly complicated process, something journals like require. editorial guidelines request researchers share code unless reports software, (4). these specify what type shared, nor reviewers encouraged verify produces article. As result, reusability shared limited. GitHub, repository does include weights; therefore, cannot process images. made together weights instructions how apply My group working comparing start extensive several tools contribute global fight against 2019.Disclosures Conflicts Interest: B.v.G. Activities article: disclosed no relevant relationships. institution received grants Delft Imaging, Siemens Healthineers, MeVis Solutions; royalties owns Thirona stock. Other relationships: relationships.References1. Rubin GD, Ryerson CJ, Haramati LB, al. Role Imaging Patient Management Pandemic: Multinational Consensus Statement Society. 2020;296(1):172–180. Link, Scholar2. Wong HYF, Lam HYS, Fong AHT, Frequency Distribution Findings Patients Positive COVID-19. 2020;296(2):E72–E78. Scholar3. RM, Sheng J, Dutta S, DeepCOVID-XR: An Algorithm Detect Trained Tested Large U.S. Clinical Data Set. 2021;299:E167–E176. Scholar4. K, Smits H, Knoops AJG, Radiographs: Multireader Evaluation System. 2020;296(3):E166–E172. Scholar5. R, Tie X, Qi Z, Diagnosis Pneumonia Using Radiography: Value Intelligence. 10.1148/radiol.2020202944. online 24, Scholar6. Baltruschat IM, Steinmeister L, Ittrich When Does Bone Suppression Lung Field Segmentation Improve X-Ray Disease Classification? In: IEEE 16th International Symposium Biomedical (ISBI 2019), Venice, Italy, April 8–11, Piscataway, NJ: IEEE, 2019; 1362–1366. Crossref, Scholar7. Kurstjens der Horst A, Herpers Rapid identification SARS-CoV-2-infected emergency department routine testing. Clin Chem Lab Med 2020;58(9):1587–1593. Medline, Scholar8. Li MD, Arun NT, Gidwani M, Automated Assessment Tracking Pulmonary Severity Convolutional Siamese Neural Networks. Radiol Artif Intell 2020;2(4):e200079 https://doi.org/10.1148/ryai.2020200079. ScholarArticle HistoryReceived: Nov 8 2020Revision requested: 16 received: 2020Accepted: 2020Published online: print: Apr 2021 FiguresReferencesRelatedDetailsCited ByCoronavirus cough sounds: approachesKazemAskari Nasab, JamalMirzaei, AlirezaZali, SarfenazGholizadeh, MeisamAkhlaghdoust2023 | Frontiers Intelligence, Vol. 6Evaluation clinical AI-based application X-raysJulius HenningNiehoff, JanaKalaitzidis, Jan RobertKroeger, DeniseSchoenbeck, JanBorggrefe, Arwed EliasMichael2023 Scientific Reports, 13, 1The plain interpretation Covid-19 pandemicDanaAlNuaimi, ReemAlKetbi2022 BJR|Open, 4, 1AI We Future Light Ouroboros Model: Plea PluralityKnudThomsen2022 AI, 3, 4COVID-19 X-ray Classification Transformer NetworksTuanLe Dinh, Suk-HwanLee, Seong-GeunKwon, Ki-RyongKwon2022 Applied Sciences, 12, 10Advances Technologies PracticeJashasmitaPal, SubhalaxmiDas2022Long-COVID diagnosis: From AI-driven modelsRiccardoCau, GavinoFaa, ValentinaNardi, AntonellaBalestrieri, JosepPuig, Jasjit SSuri, RobertoSanFilippo, LucaSaba2022 European Journal 148Development prospective screening model attending departmentsIgnatDrozdov, BenjaminSzubert, ElainaReda, PeterMakary, DanielForbes, Sau LeeChang, AbinayaEzhil, SrikanthPuttagunta, MarkHall, ChrisCarlin, David J.Lowe2021 11, 1Identification Applying Intelligent Computational Framework Based Deep Learning Machine TechniquesYarMuhammad, Mohammad DahmanAlshehri, Wael MohammedAlenazy, TruongVinh Hoang, RyanAlturki, MuhammadUsman2021 Mobile Information Systems, 2021Lessons Pandemic Use Digital Submission Survey Investigate Opinion InsidersDanieleGiansanti, IvanoRossi, LisaMonoscalco2021 Healthcare, 9, 3Application COVID-19Eun YoungKim, Myung JinChung2021 Korean Association, 64, 10Accompanying ArticleDeepCOVID-XR: SetNov 2020RadiologyRecommended Articles RSNA Education Exhibits Case Collection Metrics Altmetric Score Open AccessThis via PMC Access Subset unrestricted re-use analyses any form means acknowledgement source. permissions granted duration until revoked writing. Upon expiration permissions, perpetual license make Europe PMC, existing copyright protections.PDF download