Deep Learning Improves Predictions of the Need for Total Knee Replacement
2020; Radiological Society of North America; Volume: 296; Issue: 3 Linguagem: Inglês
10.1148/radiol.2020202332
ISSN1527-1315
Autores Tópico(s)Orthopedic Infections and Treatments
ResumoHomeRadiologyVol. 296, No. 3 PreviousNext Reviews and CommentaryFree AccessEditorialDeep Learning Improves Predictions of the Need for Total Knee ReplacementMichael L. Richardson Michael L. Richardson Author AffiliationsFrom the Department of Radiology, University of Washington, 4245 Roosevelt Way NE, Seattle, WA 98105.Address correspondence to the author (e-mail: [email protected]).Michael L. Richardson Published Online:Jun 23 2020https://doi.org/10.1148/radiol.2020202332MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Leung et al in this issue.Dr Michael Richardson is a professor of radiology and orthopedic surgery in the musculoskeletal section of the Department of Radiology at the University of Washington in Seattle. His research interests focus on radiology education, musculoskeletal imaging, and applications of computers, biostatistics, numerical analysis, and artificial intelligence in radiology. Dr Richardson is a fellow of the Association of University Radiologists and plays the fiddle, guitar, banjo, mandolin, and harmonica for folk dancing.Download as PowerPointOpen in Image Viewer Diagnosis of osteoarthritis (OA) is not difficult. A 1st-year medical student can accurately diagnose OA on a knee radiograph if the findings are sufficiently advanced. Rather, the challenge in imaging OA lies in estimating how bad it is, and whether a total knee replacement (TKR) is warranted. Systems such as the Kellgren-Lawrence (KL) grade (1), the Osteoarthritis Research Society International atlas (2), and the American Academy of Orthopedic Surgeons Clinical Practice Guidelines (3) have been used for many years to make these determinations. Once a particular grading system for the severity of OA has been chosen from among the many contenders and their variants, there is then the matter of applying that system consistently among different patients and different health care providers.The main goal of the study by Leung et al (4) in this issue of Radiology was to develop an automated system to reliably predict which patients with OA would most likely progress to needing a TKR. Previous studies that used deep learning to assess OA of the knee focused on classifying knee radiographs into one of five KL grades (5,6). The rationale was that these KL grades could then be used to predict a patient’s need for TKR. Indeed, the deep learning system designed and trained by the authors correctly classified knee radiographs into KL grades at a level comparable to that of human graders.However, the deep learning model of Leung et al (4) goes a step further, cuts out the middleman (KL classification), and predicts TKR need directly from the patients’ knee radiographs. As the authors have shown, their model predicts TKR need better than models on the basis of the conventional KL or Osteoarthritis Research Society International grading systems. One may speculate as to the reason for this improved performance. A reason might be that these conventional grading systems take the millions of pixels in a knee radiograph and compress them down to a single number for the KL system or six numbers and three features for the Osteoarthritis Research Society International system. It is likely that the radiographs contain useful diagnostic and prognostic information that is not captured by these extremely compressed grading systems. As an analogy, a meteorologist who uses all of the information from a sky photograph, such as cloud shape, size, position, altitude and other image features, would do a much better job at weather prediction than a meteorologist who relies solely on a 0–5 sky grade.Deep learning models that use convolutional neural networks are becoming increasingly more sophisticated at spotting subtle patterns in images (7). One of the engaging aspects of the convolutional neural networks used in deep learning is that a human expert does not need to predefine which imaging features the deep learning model should use to make its diagnoses or predictions. Rather, the deep learning model discovers these factors on its own as it is trained on large numbers of preclassified images.Once a convolutional neural network-based model has been trained, one would like to know which features the model used to make its predictions. Unfortunately, determining this can often be somewhat opaque. In this study, Leung et al (4) did not identify such features. However, the Gradient-weighted Class Activation Mapping, or Grad-CAM, heatmaps they derived suggested that their model concentrated on the same areas of the image that a radiologist would examine (ie, the joint spaces, the intercondylar notch of the femur, and the tibial spines).The limitations of this study are outlined nicely by the authors and point out opportunities for future investigation. For example, the study only considered anteroposterior radiographs in making its predictions. It is possible that adding information from other radiographic views and from other imaging methods such as CT, MRI, US, and nuclear medicine might further refine their model and improve its predictive ability (8).It would also be helpful to train the model on a much larger set of knee images. This would almost certainly improve the predictive performance. A larger data set could also be used to deal with the censored (incompletely known) data seen in this study and in other long-term follow-up studies. For many practical reasons, most follow-up studies are terminated at some finite number of years (9 years, in this case), even though many patients have not yet reached the study end point (TKR, in this case). Sufficiently large data sets could give investigators the statistical power to use techniques such as survival or hazard analysis to not only correct for such censored data but also to exploit time-dependent information to predict the risk of TKR and also the actual time to TKR.Leung et al (4) chose TKR as their study end point and trained their model to predict the need for TKR. However, there are other clinically relevant end points that occur after TKR. It would therefore be useful to develop a deep learning model that could accurately predict outcomes such as pain scores, physical function, and quality of life after TKR.Finally, models such as this need to be validated on an external data set. Recent experience has shown that deep learning models may not generalize to new data as well as was previously believed. External validation can reveal a lack of generalization, spot evidence of model overfitting, and recognize when deep learning models focus on spurious radiographic findings rather than actual pathologic findings (9).The availability of an automated system that could accurately predict TKR need has important and practical ramifications. An estimated 14 million Americans have symptomatic OA of the knee (10). There is therefore a great necessity to begin disease-modifying therapies as soon as possible to prevent or delay the need for a TKR. The system described by Leung et al (4) is a hopeful step in that direction.Disclosures of Conflicts of Interest: M.L.R. disclosed no relevant relationships.References1. Kellgren JH, Lawrence JS. Radiological assessment of osteo-arthrosis. Ann Rheum Dis 1957;16(4):494–502. Crossref, Medline, Google Scholar2. Altman R, Asch E, Bloch D, et al. Development of criteria for the classification and reporting of osteoarthritis. Classification of osteoarthritis of the knee. Diagnostic and Therapeutic Criteria Committee of the American Rheumatism Association. Arthritis Rheum 1986;29(8):1039–1049. Crossref, Medline, Google Scholar3. Quinn RH, Murray JN, Pezold R, Sevarino KS. Surgical management of osteoarthritis of the knee. J Am Acad Orthop Surg 2018;26(9):e191–e193. Crossref, Medline, Google Scholar4. Leung K, Zhang B, Tan J, et al. Prediction of total knee replacement and diagnosis of osteoarthritis by using deep learning on knee radiographs: data from the osteoarthritis initiative. Radiology 2020;296:584–593. Link, Google Scholar5. Abedin J, Antony J, McGuinness K, et al. Predicting knee osteoarthritis severity: comparative modeling based on patient’s data and plain X-ray images. Sci Rep 2019;9(1):5761. Crossref, Medline, Google Scholar6. Tiulpin A, Thevenot J, Rahtu E, Lehenkari P, Saarakkala S. Automatic knee osteoarthritis diagnosis from plain radiographs: A deep learning-based approach. Sci Rep 2018;8(1):1727. Crossref, Medline, Google Scholar7. Chartrand G, Cheng PM, Vorontsov E, et al. Deep learning: A primer for radiologists. RadioGraphics 2017;37(7):2113–2131. Link, Google Scholar8. Roemer FW, Demehri S, Omoumi P, et al. State of the Art: Imaging of Osteoarthritis-Revisited 2020. Radiology 2020. 10.1148/radiol.2020192498. Published online May 19, 2020. Link, Google Scholar9. Zech JR, Badgeley MA, Liu M, Costa AB, Titano JJ, Oermann EK. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLoS Med 2018;15(11):e1002683. Crossref, Medline, Google Scholar10. Vina ER, Kwoh CK. Epidemiology of osteoarthritis: literature update. Curr Opin Rheumatol 2018;30(2):160–167. Crossref, Medline, Google ScholarArticle HistoryReceived: May 21 2020Revision requested: June 1 2020Revision received: June 8 2020Accepted: June 8 2020Published online: June 23 2020Published in print: Sept 2020 FiguresReferencesRelatedDetailsCited ByTotal Knee ArthroplastyEmreTokgoz, SarahLevitt, DianaSosa, Nicholas A.Carola, VishalPatel2023Artificial intelligence in diagnosis of knee osteoarthritis and prediction of arthroplasty outcomes: a reviewLok SzeLee, Ping KeungChan, ChunyiWen, Wing ChiuFung, AmyCheung, Vincent Wai KwanChan, Man HongCheung, HenryFu, Chun HoiYan, Kwong YuenChiu2022 | Arthroplasty, Vol. 4, No. 1Toward Generalizability in the Deployment of Artificial Intelligence in Radiology: Role of Computation Stress Testing to Overcome UnderspecificationThomas Eche, Lawrence H. Schwartz, Fatima-Zohra Mokrane, Laurent Dercle, 27 October 2021 | Radiology: Artificial Intelligence, Vol. 3, No. 6Accompanying This ArticlePrediction of Total Knee Replacement and Diagnosis of Osteoarthritis by Using Deep Learning on Knee Radiographs: Data from the Osteoarthritis InitiativeJun 23 2020RadiologyRecommended Articles Prediction of Total Knee Replacement and Diagnosis of Osteoarthritis by Using Deep Learning on Knee Radiographs: Data from the Osteoarthritis InitiativeRadiology2020Volume: 296Issue: 3pp. 584-593Are Artificial Intelligence Challenges Becoming Radiology’s New “Bee’s Knees”?Radiology: Artificial Intelligence2021Volume: 3Issue: 3Development and Validation of a Multitask Deep Learning Model for Severity Grading of Hip Osteoarthritis Features on RadiographsRadiology2020Volume: 295Issue: 1pp. 136-145Automatic Deep Learning–assisted Detection and Grading of Abnormalities in Knee MRI StudiesRadiology: Artificial Intelligence2021Volume: 3Issue: 3Radiology Alchemy: GAN We Do It?Radiology: Artificial Intelligence2021Volume: 3Issue: 5See More RSNA Education Exhibits Structured Reporting And Natural Language Processing: A Successful MarriageDigital Posters2021MRI of the Metabolic KneeDigital Posters2022Artificial Intelligence in Diagnostic Imaging: Current Applications and Future PerspectiveDigital Posters2019 RSNA Case Collection Primary Bone LymphomaRSNA Case Collection2020Cerebral Fat EmbolismRSNA Case Collection2020Osteoid OsteomaRSNA Case Collection2020 Vol. 296, No. 3 Metrics Altmetric Score PDF download
Referência(s)