Session: (2387–2424) Vasculitis – Non-ANCA-Associated & Related Disorders Poster III
2424: Exploring the Limit of Image Resolution for Human Expert Classification of Vascular Ultrasound Images in Giant Cell Arteritis and Healthy Subjects: The GCA-US-AI Project
Clinic of Internal Medicine III, Department of Oncology, Hematology, Rheumatology and Clinical Immunology, University Hospital Bonn Bonn, Germany
Disclosure information not submitted.
Valentin Sebastian Schäfer1, Stavros Chrysidis2, Christian Dejaco3, Sara Monti4, Matthew Koster5, Pantelis Karakostas1, Wolfgang Schmidt6, Eugenio De Miguel7, Christina Duftner8, Alojzija Hocevar9, Annamaria Iagnocco10, Marcin Milchert11, Chetan Mukhtyar12, Cristina Ponte13, Lene Terslev14, Tanaz Kermani15, Uffe Møller Døhn16, Berit Dalsgaard Nielsen17, Aaron Juche18, Luca Seitz19, Minna Kohler20, Kresten Keller21, Rositsa Karalilova22, Thomas Daikeler23, Sarah Mackie24, Karina Torralba25, Kornelis van der Geest26, Dennis Boumans27, Philipp Bosch28, Alessandro Tomelleri29, Markus Aschwanden30, Peter Brossart1, Charlotte Behning31 and Claus Juergen Bauer1, 1Clinic of Internal Medicine III, Department of Oncology, Hematology, Rheumatology and Clinical Immunology, University Hospital of Bonn, Bonn, Germany, 2Department of Rheumatology, Southwest Jutland Hospital Esbjerg, Esbjerg, Denmark, 3Department of Rheumatology, Medical University Graz, Graz, Austria; Department of Rheumatology, Hospital of Bruneck (ASAA-SABES), Teaching Hospital of the Paracelsius Medical University, Brunico, Italy, 4Division of Rheumatology, Fondazione IRCCS Policlinico San Matteo, Pavia, Italy; Department of Internal Medicine and Therapeutics, Università di Pavia, Pavia, Italy, 5Mayo Clinic, Rochester, MN, 6Rheumatology, Immanuel Krankenhaus Berlin, Medical Centre for Rheumatology Berlin-Buch, Berlin, Germany, 7Hospital Universitario La Paz, Madrid, Spain, 8Department of Internal Medicine, Clinical Division of Internal Medicine II, Medical University Innsbruck, Innsbruck, Austria, 9Department of Rheumatology, Universitiy Medical Centre Ljubljana, Ljubljana, Slovenia, 10University of Turin, Roma, Italy, 11Department of Internal Medicine, Rheumatology, Diabetology, Geriatrics and Clinical Immunology, Pomeranian Medical University in Szczecin, Szczecin, Poland, 12Vasculitis service, Rheumatology department, Norfolk and Norwich University Hospital, Norwich, United Kingdom, 13Department of Rheumatology, Centro Hospitalar Universitário Lisboa Norte, Centro Académico de Medicina de Lisboa, Lisbon, Portugal; Rheumatology Research Unit, Instituto de Medicina Molecular, Faculdade de Medicina, Universidade de Lisboa, Centro Académico de Medicina de Lisboa, Lisbon, Portugal, 14Center for Rheumatology and Spine Diseases, Rigshospitalet, Glostrup, Denmark, 15Rheumatology, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, 16Copenhagen Center for Arthritis Research (COPECARE), Center for Rheumatology and Spine Diseases, Rigshospitalet, Copenhagen, Denmark, 17Department of Rheumatology, Aarhus University Hospital, Aarhus, Denmark; Department of Medicine, The Regional Hospital in Horsens, Horsens, Denmark, 18Department of Rheumatology, Immanuel Hospital, Berlin, Germany, 19Rheumatology and Immunology, Inselspital University Hospital Bern, Bern, Switzerland, 20Massachusetts General Hospital, Harvard Medical School, Boston, MA, 21Department of Rheumatology, Aarhus University Hospital; Department of Clinical Medicine, Aarhus University, Aarhus, Denmark, 22Clinic of Rheumatology, Medical University Plovdiv, Plovdiv, Bulgaria, 23Clinic for Rheumatology, University Hospital Basel, Basel, Switzerland, 24Leeds Institute of Rheumatic and Musculoskeletal Medicine, University of Leeds, Leeds, UK; Leeds Biomedical Research Centre, Leeds Teaching Hospitals NHS Trust, Leeds, United Kingdom, 25Division of Rheumatology, Department of Medicine, Loma Linda University School of Medicine, Loma Linda, CA, 26Rheumatology and Clinical Immunology, University Medical Center Groningen, University of Groningen, Groningen, Netherlands, 27Rheumatology and Clinical Immunology, Hospital Group Twente, Almelo, Netherlands, 28Medical University of Graz, Graz, Austria, 29Unit of Immunology, Rheumatology, Allergy and Rare Diseases, San Raffaele Scientific Institute, Milano, Italy, 30Department of Angiology, University Hospital Basel, Basel, Switzerland, 31Institute of Medical Biometry, Informatics and Epidemiology, University Hospital of Bonn, Bonn, Germany
Background/Purpose: Giant cell arteritis (GCA) is the most common form of vasculitis in adults, necessitating prompt diagnosis to prevent severe complications. However, access to expert GCA ultrasonography is often limited to larger medical centers. The utilization of artificial intelligence (AI)-based assistance for ultrasound image classification holds significant potential. The development of such AI systems involves selecting a neural network architecture from a wide range of options, each with distinct characteristics, including limitations in processing image size, resolution, result quality, and resource efficiency. This project aims to address the minimum resolution required for human experts to reliably classify ultrasound images of commonly affected arteries, distinguishing between GCA and normal.
Methods: In this study, a blinded classification task was conducted, involving 42 international experts from the OMERACT subgroup on ultrasound in large vessel vasculitis. A set of 250 vascular ultrasound images from both GCA patients and healthy individuals were presented to the experts. The image selection process initially comprised 10 B-mode scans each of the common temporal artery, its frontal branch, parietal branch, and the axillary artery (all in cross-section), as well as the axillary artery in longitudinal scan. These 50 images were presented in a random order at five distinct resolution levels: 32x32, 64x64, 128x128, 224x224, and 512x512 pixels (refer to Figure 1). The survey and image classification tasks were conducted using REDCap version 13.2.5, while data interim analysis was performed utilizing R (version 4.3.0).
Results: This interim analysis includes data from 30 study group members with an average of 10.5±6.4 years (mean±standard deviation) of experience in ultrasound in GCA. Across all artery categories, the proportion of images considered unclassifiable decreased as the image resolution increased. The mean percentage of unclassifiable images was 91% at 32x32 pixels and reduced to 26% at 512x512pixels (refer to Table 1). Similarly, lower image resolution resulted in less frequent correct classifications (see Figure 2). The chance of correct image classification was comparable at image resolutions of 512x512 pixels (highest resolution level) and 224x224 pixels (odds ratio = 0.99, p=0.891), while it was significantly lower at image sizes of 128x128 pixels (OR=0.79, p=0.01), 64x64 pixels (OR = 0.42, p< 0.001), and 32x32 pixels (OR=0.24, p< 0.001).
Conclusion: As the image resolution increases, the proficiency of human experts in GCA ultrasound improves, enabling more confident and accurate classification of vascular ultrasound images. At a resolution of 224x224 pixels and above, classification accuracy exceeded 50%. The best classification performance was achieved at a resolution of 512x512 pixels (74.13% of all single-image classifications correct). Yet given the comparable performance at a resolution level of 224x224 pixels (73.93% of all single-image classifications correct), this opens up a variety of perspectives in neural network architecture selection for the development of an AI based assistance for GCA ultrasound image classification.
Figure 1: Shown is an exemplary ultrasound image of the common temporal artery (cross-section, B-mode), presented to study participants at five different resolution levels.
Table 1: Frequency of the selected response category "Meaningful classification is not possible" depending on the image resolution level.
Figure 2: Frequency of correct image classification as "GCA patient" or "healthy individual" depending on image resolution level.
V. Schäfer: None; S. Chrysidis: None; C. Dejaco: AbbVie/Abbott, 1, 5, 6, Eli Lilly, 6, Galapagos Pharma, 1, 6, Janssen, 1, 6, Novartis, 1, 6, Pfizer, 6, Sparrow, 1; S. Monti: CSL Vifor, 6; M. Koster: None; P. Karakostas: None; W. Schmidt: AbbVie/Abbott, 1, 5, 6, Amgen, 1, 6, Bristol-Myers Squibb(BMS), 6, Chugai, 6, GlaxoSmithKlein(GSK), 1, 5, 6, Janssen, 6, Medac, 6, Novartis, 1, 5, 6, Roche, 6, UCB, 6; E. De Miguel: None; C. Duftner: None; A. Hocevar: None; A. Iagnocco: None; M. Milchert: None; C. Mukhtyar: None; C. Ponte: None; L. Terslev: Eli Lilly, 1, Janssen, 1, 6, Novartis, 6, UCB, 1; T. Kermani: None; U. Møller Døhn: None; B. Nielsen: None; A. Juche: None; L. Seitz: None; M. Kohler: Mymee, 2, Springer Publications, 9; K. Keller: None; R. Karalilova: None; T. Daikeler: None; S. Mackie: AbbVie/Abbott, 2, AstraZeneca, 2, GlaxoSmithKlein(GSK), 3, 12, Investigator, National Institute for Health and Care Research, 5, 12, investigator on STERLING-PMR trial, funded by NIHR; patron of the charity PMRGCAuk, Pfizer, 2, 6, Roche, 2, 6, 12, Support from Roche/Chugai to attend EULAR2019 in person, Sanofi, 2, 12, Investigator, Sparrow, 12, Investigator, UCB and Novartis, 6, Vifor, 6; K. Torralba: None; K. van der Geest: None; D. Boumans: None; P. Bosch: Janssen, 6, 12, Congress fees, Pfizer, 5; A. Tomelleri: Novartis, 1; M. Aschwanden: None; P. Brossart: None; C. Behning: None; C. Bauer: None.