摘要
HomeRadiologyVol. 301, No. 2 PreviousNext Reviews and CommentaryFree AccessEditorialAn Artificially Intelligent Solution for a Real Problem in Musculoskeletal Radiology: Bone TumorsJohn A. Carrino John A. Carrino Author AffiliationsFrom the Department of Radiology and Imaging, Weill Cornell Medicine, Hospital for Special Surgery, 535 E 70th St, 3E-012, New York, NY 10021.Address correspondence to the author (e-mail: [email protected]).John A. Carrino Published Online:Sep 7 2021https://doi.org/10.1148/radiol.2021211560MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by von Schacky et al in this issue.Dr Carrino is a musculoskeletal radiologist and the vice chair of radiology and imaging at the Hospital for Special Surgery in New York, NY, and is a professor of radiology at Weill Cornell Medicine. He has authored more than 250 peer-reviewed publications. Research interests focus on the spine, rheumatology, metabolic bone diseases, and neuromuscular disorders. He serves on NIH, DoD, and RSNA study sections and participates in the Digital Communications in Medicine standards committee.Download as PowerPointOpen in Image Viewer Radiography has long been the first-line imaging technique used to assess bone lesions. Tumor and tumorlike lesions can be particularly problematic for general and subspecialty radiologists of all experience levels. The lesions are uncommon and are therefore often unfamiliar. The differential diagnosis is broad. A few lesions have a characteristic appearance, while others are nonspecific. The conspicuity can range from markedly destructive blowout lesions to subtle regions of trabecular rarefaction. However, a key issue is to discriminate benignancy from malignancy.Radiographic evaluation is still critical in the evaluation of bone tumors. Newer techniques also play an important role in characterization and treatment of these neoplasms. Although advanced imaging modalities such as MRI can help confirm the presence of a lesion, map the extent, and assess treatment response, diagnosis usually relies on radiographic characterization (1). Pathologic examination of bone is also heavily reliant on radiographic imaging, and many expert musculoskeletal pathologists would be reluctant to interpret tissue samples without a radiograph.The musculoskeletal radiologist’s approach to the radiographic diagnosis of bone tumors consists of analyzing the lesion in an organized fashion, with attention to the specific features of tumor location, margins, and zone of transition; periosteal reaction; mineralization; size and number of lesions; and presence of a soft-tissue component (2). Patient age is also an important clinical factor in the diagnosis of bone tumors because various lesions have predilections for specific age groups.Radiography remains the most important imaging test in the diagnosis of bone tumors and is considered highly appropriate by the American College of Radiology Appropriateness Criteria (3). In current clinical practice, bone lesions may be inappropriately characterized by confounding benign and malignant entities. Many musculoskeletal radiographs may not be interpreted by subspeciality radiologists but rather by general radiologists, orthopedic surgeons, or emergency physicians. Therefore, an aid to detect and classify lesions even at the basic level of benign or malignant would facilitate workflow and clinical care.In this issue of Radiology, von Schacky and colleagues (4) describe experiments on the ability of a deep learning system to differentiate primary malignant from benign bone tumors with impressive results. The authors were motivated by a relevant clinical problem and designed a study to leverage machine learning to address it. The accuracy of the system was better than that of radiology residents and equivalent to that of experienced radiologists. This report extends previous work in this domain, as the number of examinations and the results are similar to or better than those in the recent literature. Overall, this is a well-executed study: the authors followed the guide on assessing radiology research on artificial intelligence (AI) to a large extent and used many of the key considerations provided by the Radiology editorial board (5). All three image sets (training, testing, validation) were well defined and independent from each other without overlap. The test set was from an external institution and used for final statistical assessment. The authors provided justification for the sample size of each data set. The AI algorithm used a robust reference standard of pathologic examination. The image preparation is well described. The AI performance was benchmarked against the performance of experienced subspecialty radiologists.The authors proposed an automated technique to classify, detect, and segment bone tumors on radiographs. Images from different sources were combined to train the algorithm, a definitive standard of truth (pathologic examination) was available, the algorithm and the tools used to generate it were fully disclosed, and the performance of the algorithm was compared with that of radiology experts. Nonetheless, the methods and results raise several interesting items. The added value of having both detection and segmentation output in the clinical workflow is somewhat controversial (as only one of them would already provide localization). The authors used a multitask architecture to improve model performance, but the insight as to how the deep learning system made its decisions is lost in this process. The data sets used all had a lesion present, and the lack of normal radiographs does not inform on the true detection ability or specificity (ie, “normal in health” as sometimes one of the hardest tasks is just to identify a radiograph as normal). This tool was tested only with cases that went for histopathologic examination, which potentially introduced a spectrum bias. Performance may drop, perhaps substantially, in real-world use. However, it may be exactly this group of more indeterminate cases necessitating biopsy where a tool like this would be more useful. Future avenues of research should look at multiple lesions and should include lesions other than primary bone lesions in large data sets where the majority of images will not have a bone lesion to better mimic the real world.In summary, this study is an important step forward in the progress of useful AI in practice for the detection and characterization of bone lesions for these often incidentally encountered lesions in general practice, musculoskeletal medicine clinics (orthopedics, rheumatology), and emergency departments. This type of AI algorithm could easily be integrated into patient care, operating in the background as a preprocessing function, or invoked at the time an observer identifies a lesion. Open questions include whether multiple lesions of potentially different origins can be detected and accurately characterized. Although AI offers information with regard to aggressiveness of the lesion, it likely will not obviate the need for other imaging modalities, such as MRI (6) or PET/CT, and it is not likely to preclude biopsy for definitive diagnosis.Disclosures of Conflicts of Interest: J.A.C. is a member of the Radiology editorial board.References1. Mintz DN, Hwang S. Bone tumor imaging, then and now: review article. HSS J 2014;10(3):230–239. Crossref, Medline, Google Scholar2. Miller TT. Bone tumors and tumorlike conditions: analysis with conventional radiography. Radiology 2008;246(3):662–674. Link, Google Scholar3. Berquist TH, Dalinka MK, Alazraki N, et al. Bone tumors. American college of radiology ACR appropriateness criteria. Radiology 2000;215(Suppl):261–264. Medline, Google Scholar4. von Schacky CE, Wilhelm NJ, Schafer VS, et al. Multitask deep learning for segmentation, and classification of primary bone tumors on radiographs. Radiology 2021.https://doi.org/10.1148/radiol.2021204531. Published online September 7, 2021. Link, Google Scholar5. Bluemke DA, Moy L, Bredella MA, et al. Assessing Radiology Research on Artificial Intelligence: A Brief Guide for Authors, Reviewers, and Readers—From the Radiology Editorial Board. Radiology 2020;294(3):487–489. Link, Google Scholar6. Fayad LM, Jacobs MA, Wang X, Carrino JA, Bluemke DA. Musculoskeletal tumors: how to use anatomic, functional, and metabolic MR techniques. Radiology 2012;265(2):340–356. Link, Google ScholarArticle HistoryReceived: June 21 2021Revision requested: July 2 2021Revision received: Aug 4 2021Accepted: Aug 9 2021Published online: Sept 07 2021Published in print: Nov 2021 FiguresReferencesRelatedDetailsAccompanying This ArticleMultitask Deep Learning for Segmentation and Classification of Primary Bone Tumors on RadiographsSep 7 2021RadiologyRecommended Articles Imaging Findings of Metabolic Bone DiseaseRadioGraphics2016Volume: 36Issue: 6pp. 1871-1887Common Skeletal Neoplasms and Nonneoplastic Lesions at 18F-FDG PET/CTRadioGraphics2021Volume: 42Issue: 1pp. 250-267Periosteal Pathologic Conditions: Imaging Findings and PathophysiologyRadioGraphics2022Volume: 43Issue: 2Tumor-induced Osteomalacia Secondary to Phosphaturic Mesenchymal TumorRadiology: Imaging Cancer2023Volume: 5Issue: 2Craniofacial Manifestations of Systemic Disorders: CT and MR Imaging Findings and Imaging ApproachRadioGraphics2018Volume: 38Issue: 3pp. 890-911See More RSNA Education Exhibits For Every Bone Aggression, There is a Periosteal ReactionDigital Posters2019The Role of Conventional Radiography in the Diagnosis of Bone TumorsDigital Posters2022Through Thick and Thin: Periosteal Reactions and Their Underlying EtiologyDigital Posters2019 RSNA Case Collection Osteoid osteoma of the femurRSNA Case Collection2022Chronic Recurrent Multifocal OsteomyelitisRSNA Case Collection2021Chondroblastoma of the glenoidRSNA Case Collection2021 Vol. 301, No. 2 Metrics Altmetric Score PDF download