Isophotes, Scale Space, and Invariants in Lung CT for COPD Diagnosis
2022; Radiological Society of North America; Volume: 4; Issue: 1 Linguagem: Inglês
10.1148/ryai.210301
ISSN2638-6100
Autores Tópico(s)Chronic Obstructive Pulmonary Disease (COPD) Research
ResumoHomeRadiology: Artificial IntelligenceVol. 4, No. 1 Previous CommentaryFree AccessIsophotes, Scale Space, and Invariants in Lung CT for COPD DiagnosisMichael W. Vannier Michael W. Vannier Author AffiliationsFrom the Department of Radiology, University of Chicago Medical Center, 5841 S Maryland Ave, Chicago, IL 60637.Address correspondence to the author (e-mail: [email protected]).Michael W. Vannier Published Online:Jan 19 2022https://doi.org/10.1148/ryai.210301MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked InEmail See also the article by Savadjiev et al in this issue.Michael W. Vannier, MD, is a professor emeritus of radiology at the University of Chicago and was professor of radiology at the University of Iowa and special assistant to the director of biomedical imaging at the National Cancer Institute. He has been editor-in-chief of various journals, president of the International Society for Computer Aided Surgery, and the Georgia Eminent Scholar in medical imaging. He has served as a professor and vice-chairman for research at Mallinckrodt Institute of Radiology and was inducted into the U.S. Space Foundation Hall of Fame for pioneering work in digital medical imaging and multispectral analysis.Download as PowerPointOpen in Image Viewer In 1966, Gwilym S. Lodwick, MD, an academic radiologist from the University of Missouri, was among the first to report on computer-aided diagnosis of chest disease (1) and was awarded a National Institutes of Health research grant to pursue this work. Thus began 50 years of computer analysis in chest imaging (2) that serve as a prelude to a new landmark study of chronic obstructive pulmonary disease (COPD) diagnosis using artificial intelligence (3). This study by Savadjiev et al in this issue of Radiology: Artificial Intelligence introduces a novel metric, the mean curvature of isophotes (MCI), to characterize lung structure. The authors draw on extensive experience in computer vision, astronomy, and other fields with the development of mathematical tools that extract deep structure from CT images. The results are features that are robust and invariant to the influence of grayscale, spatial, and rotational effects.Biologic structures possess details at all levels of scale, including ultra-, micro-, and macrostructural features. Widely used feature extraction methods, such as those frequently employed for medical image processing and segmentation in visualization and radiomics, are typically sensitive to scale (or magnification) changes, grayscale, and rotation, which can lead to uncontrolled variability. The widely used convolutional neural networks (CNNs) developed for medical imaging have almost universally been constructed on the basis of these features despite their intrinsic limitations.Isophotes are widely used in astronomy to delineate galaxies as ellipses contained within a boundary of uniform intensity light. In CT imaging, we can observe “isophotes” by narrowing the display window width to zero and varying the window level. Over a range of common CT attenuation values, one or more closed contours will appear in most images. These isoattenuation contours are isophotes that can be used to quantify the two-dimensional distribution of light, for example, in galaxies for astronomical photographs or anatomic structures in CT images. The geometry of isophotes that are essential for localization, measurement, and characterization of galaxies exist at multiple scales, akin to biologic structures. Software tools for isophote extraction and measurement are incorporated in programming libraries for astronomy and astrophysics such as Astropy. The Astropy Project is a community effort to develop a common core package for astronomy in the Python language and foster an ecosystem of interoperable astronomy packages (https://www.astropy.org). Geometric properties of isophotes are essential elements in image analysis techniques, due to their invariance under general invertible intensity transformations (4). Isophote curvature invariance yields important benefits when dealing with multiscale and multiresolution images often found with clinical CT scans where the structure is complex while acquisition and reconstruction parameters are not fixed a priori.The Savadjiev et al article (3) is a composite of a technical report introducing a new method for computation of isophotes and related image measures invariant to gray-level intensity transformations (Appendices E1–E8) that were tested in a clinical CT study. Their isophote-based technique leads to an efficient tool for extraction of MCI found in CT scans that is robust in scale space. This supplement stands alone as a major contribution to CT image analysis, supported by 27 references from publications in computer vision, astrophysics, computer graphics, and radiomics. In addition to the technical report on MCI, the main body of the article contains a multi-institutional clinical feasibility study where chest CT scans were compared with global initiative for chronic obstructive lung disease (or, GOLD) spirometry classification, demonstrating significantly better performance than conventional CNNs such as ResNet (5). The authors point out that the lack of a well-defined relationship between spirometry-based ground truth staging and radiologic appearance makes the automated image-based detection and staging of COPD challenging, even for powerful algorithms such as CNNs.The notion of scale space is supported by a large body of applied mathematics pioneered by Koenderink (6), ter Haar Romeny (4), Lindeberg (7), and their associates and students over the past 3 decades, revolutionizing computer vision and yielding many tools and techniques that support a broad range of applications. The need for multiscale representation of image data and multiscale feature detection is widely understood and applied. These methods bear a close relationship to human vision and may provide a means to augment computer-based feature extraction that surpasses conventional features used to train and implement CNNs. The first stage of the human visual perception system has a large set of filter banks, which operate at multiple scales, where multiple orientations are present. Rigorous mathematical modeling of the human visual system to handle CT images with a practical and robust algorithm is an important achievement, as demonstrated by success with COPD classification.This study of COPD opens many avenues for future work, including that focused on other diffuse lung diseases, and may provide sensitive biomarkers to measure lung disease progression and monitor therapy. Future work applying MCI to lung and other disease states is potentiated by the authors’ contribution of source code to implement this tool on GitHub, an open software repository (http://www.github.com/). Access to their source code provides the opportunity for further innovation and independent testing in laboratories worldwide. Well-known software libraries that augment the Python computer language, such as Pytorch (http://pytorch.org/) with pretraining on the ImageNet database (http://www.image-net.org/), were used in this study. As a result, the essential elements of the software development in this project were aided by open software.To explore the potential of isophotes, scale space, and invariants for CT image analysis, it may be necessary to look beyond the medical literature. A basic understanding of photometry in astronomy can be helpful, and the U.K. Gaia satellite website provides a starting place (https://www.gaia.ac.uk/sites/default/files/resources/Photometry_in_Astronomy.pdf). Photutils is an open-source Python package that provides tools to detect astronomical sources and perform photometry (https://photutils.readthedocs.io/en/stable/). An introduction to scale-space theory and multiscale geometric image analysis that explains invariance properties is available online (https://faculty.idc.ac.il/arik/lodseminar/05ScaleSpace/romeny_scalespace.pdf).Once again, we see how radiology and chest imaging can benefit from image analysis methods that originated in other fields after they are translated by a capable group of experts who develop and test new tools that advance chest image analysis. In addition, we see the vocabulary of radiology expands to incorporate novel concepts from medical image analysis, such as isophotes, scale space, and invariants, that enrich our clinical literature.Disclosures of conflicts of interest: M.W.V. No relevant relationships.Author declared no funding for this work.References1. Lodwick GS. Computer-aided diagnosis in radiology. A research plan. Invest Radiol 1966;1(1):72–80. Crossref, Medline, Google Scholar2. van Ginneken B. Fifty years of computer analysis in chest imaging: rule-based, machine learning, deep learning. Radiol Phys Technol 2017;10(1):23–32. Crossref, Medline, Google Scholar3. Savadjiev P, Gallix B, Rezanejad M, et al. Improved Detection of Chronic Obstructive Pulmonary Disease at Chest CT Using the Mean Curvature of Isophotes. Radiol Artif Intell 2022;4(1):e210105. Google Scholar4. ter Haar Romeny BM. Multi-scale and multi-orientation medical image analysis. In: Deserno TM, ed. Biomedical Image Analysis. Berlin, Germany: Springer, 2011; 175–194. Google Scholar5. Tang LYW, Coxson HO, Lam S, Leipsic J, Tam RC, Sin DD. Towards large-scale case-finding: training and validation of residual networks for detection of chronic obstructive pulmonary disease using low-dose CT. Lancet Digit Health 2020;2(5):e259–e267. Crossref, Medline, Google Scholar6. Koenderink JJ. The structure of images. Biol Cybern 1984;50(5):363–370. Crossref, Medline, Google Scholar7. Lindeberg T. Scale-Space Theory in Computer Vision: Volume 256 of Kluwer International Series in Engineering and Computer Science. Boston, Mass: Kluwer Academic, 1994. Crossref, Google ScholarArticle HistoryReceived: Dec 6 2021Revision requested: Dec 13 2021Revision received: Dec 15 2021Accepted: Dec 20 2021Published online: Jan 19 2022 FiguresReferencesRelatedDetailsAccompanying This ArticleImproved Detection of Chronic Obstructive Pulmonary Disease at Chest CT Using the Mean Curvature of Isophotes15 Dec 2021Radiology: Artificial IntelligenceRecommended Articles RSNA Education Exhibits RSNA Case Collection Recommended Articles Vol. 4, No. 1 Metrics Altmetric Score
Referência(s)