Session Type: Poster Session A
Session Time: 8:30AM-10:30AM
Background/Purpose: Magnetic resonance imaging (MRI) plays a critical role in assisting the diagnosis of axial spondyloarthritis (axSpA). There are many characteristic MRI lesions that can be present in patients with axSpA; these have been well characterized in the literature,1 and, alongside clinical presentation and laboratory features, can be used to assist diagnosis. However, in-depth knowledge of these lesions and their definitions, as well as reliability of identification and scoring, varies amongst general radiologists and rheumatologists.2 The use of artificial intelligence to develop automated recognition of MRI lesions associated with axSpA has the potential to vastly improve diagnosis of this condition. We present a data-driven approach to automatically detect sacroiliitis on MRI.
Methods: We built a training database of de-identified sacroiliac joint (SIJ) MRI scans containing short tau inversion recovery (STIR) and T1 sequences acquired on 12 scanner models from three vendors (50% Siemens, 46% GE and 4% Philips). This dataset contains baseline MRI images from 202 patients (mean age: 38 years [range: 18–74]; 52% male) scanned at 34 clinical sites in 7 countries (97% from Europe) from interventional studies in patients with axSpA (C-axSpAnd [NCT02552212], BE MOBILE 1 [NCT03928704] and BE MOBILE 2 [NCT03928743]). 85% of the images were from patients with non-radiographic axSpA. Informed consent was collected from all patients for these studies. Three expert readers independently and manually delineated bone marrow edema (BME) lesions in a subset of cases. A positive case was defined as a patient with one or more BME lesions. We developed a two-stage lesion detection pipeline: firstly, a deep learning model binary classified each voxel present in the 3D image as either a normal or BME lesion (mimicking the reader’s delineations); secondly, this voxel-level information was aggregated to a binary case-level prediction (positive if one or more BME lesions were predicted). The time to read one patient was < 0.5 min. We then performed a stratified 7-fold cross-validation experiment comparing the automated predictions with ground truth read-outs from the readers.
Results: A total of 280 lesions in 116 BME-positive patients were identified. The presented method achieved case-level BME lesion detection results of 77% area under the receiver operating characteristic (ROC) curve (95% confidence interval [CI]: 71% to 83%; Figure 1), 74% specificity (95% CI: 64% to 83%), 70% sensitivity (95% CI: 62% to 78%) and 79% precision (95% CI: 71% to 86%).
Conclusion: Using MRI images from 202 patients, our automated method achieved robust case-level BME lesion detection results. This methodology has the potential to support radiologists and rheumatologists in detecting BME lesions on MRI scans to aid early and accurate diagnosis of axSpA. Further investigations, including confirmatory analyses and methodological improvements and validation, are warranted.
References: 1.Maksymowych WP. Ann Rheum Dis 2019; 78: 1550–58; 2. Bennett AN. J Rheumatol 2017; 44: 780–5.
To cite this abstract in AMA style:Nicolaes J, Machado P, Baraliakos X, Santosh M, Carnell A, de Peyrecave N, Bennett A. Development of a Deep Learning Algorithm for the Detection of Sacroiliitis on MRI in Patients with Active Axial Spondyloarthritis [abstract]. Arthritis Rheumatol. 2021; 73 (suppl 9). https://acrabstracts.org/abstract/development-of-a-deep-learning-algorithm-for-the-detection-of-sacroiliitis-on-mri-in-patients-with-active-axial-spondyloarthritis/. Accessed January 27, 2023.
« Back to ACR Convergence 2021
ACR Meeting Abstracts - https://acrabstracts.org/abstract/development-of-a-deep-learning-algorithm-for-the-detection-of-sacroiliitis-on-mri-in-patients-with-active-axial-spondyloarthritis/