Session Information
Date: Sunday, November 7, 2021
Session Type: Poster Session B
Session Time: 8:30AM-10:30AM
Background/Purpose: Active inflammatory changes in sacroiliac joints (SIJ) compatible with axial spondyloarthritis – axSpA (first of all, subchondral bone marrow edema – BME / osteitis) as detected by magnetic resonance imaging (MRI) play a pivotal role in diagnostic and classification approaches. Artificial intelligence/machine learning methods such as a trained artificial convolutional network (CNN) offer a potential for the development of assistant tools to be used by radiologists and rheumatologists in clinical practice.
The present study is aimed to evaluate the possibility of detection of active inflammatory changes compatible with axSpA on MRI of SIJ.
Methods: A total of 6 trained and calibrated readers evaluated MRIs of SIJ (STIR, semicoronal views) of 476 patients with and without axSpA from 4 cohorts (GESPIC-AS, GESPIC-Crohn, GESPIC-Uveitis, and OptiRef). Readers indicated the presence or absence of active inflammatory changes compatible with axSpA and specified the type of changes. Active inflammatory changes were considered present if at least 4 out of 6 readers deemed them positive. Images with an undetermined classification were adjudicated in a consensus reading session. These results were used for the training and validation of the CNN. First, we used an object detection approach to train a 3D CNN to detect the SIJ on all available MRIs to generate dilated masks of the SIJs to focus subsequent model training to the SIJs and periarticular areas. Secondly, we used a classification task to train a 2D CNN (ResNet34) to predict the presence of active inflammatory changes compatible with SpA on these grid MRI images. MRIs from the ASAS classification Cohort annotated by 7 readers (Maksymowych WP, at al. Ann Rheum Dis 2020;79:935-942) formed the holdout dataset that was used to test for CNN generalizability.
Results: A total of 403 STIR-MRIs were included in the training, 73 in the validation sets and 122 in the holdout set. The trained classification CNN achieved an accuracy of 91.8%, a sensitivity of 88.9% and a specificity of 93.5% in detecting active inflammatory changes in SIJ compatible with axSpA in the validation set (n=73). The accuracy of MRI classification in the independent holdout set (ASAS) was 81.5% with a sensitivity of 67% and a specificity of 84.5%. Figure 1 shows an exemplary class activation map of the CNN.
Conclusion: Our study shows the principal possibility of detecting active inflammatory changes compatible with axSpA on MRI of SIJs. Detection of structural changes and overall contextual interpretation of the findings are warranted for the subsequent development steps.
Acknowledgment. The project was supported by a research grant from ASAS. We want to thank colleagues who performed annotation of the images from the ASAS classification cohort: Xenofon Baraliakos, Robert Lambert, Pedro Machado, Walter Maksymowych, Mikkel Ostergaard, Suzanne Juhl Pedersen, and Ulrich Weber.
To cite this abstract in AMA style:
Poddubnyy D, Vahldiek J, Adams L, Diekhoff T, Proft F, Protopopov M, Rademacher J, Rios Rodriguez V, Torgutalp M, Niehues S, Bressem K. Artificial Neural Network for the Recognition of Active Inflammatory Changes Compatible with Axial Spondyloarthritis on Magnetic Resonance Imaging of Sacroiliac Joints [abstract]. Arthritis Rheumatol. 2021; 73 (suppl 9). https://acrabstracts.org/abstract/artificial-neural-network-for-the-recognition-of-active-inflammatory-changes-compatible-with-axial-spondyloarthritis-on-magnetic-resonance-imaging-of-sacroiliac-joints/. Accessed .« Back to ACR Convergence 2021
ACR Meeting Abstracts - https://acrabstracts.org/abstract/artificial-neural-network-for-the-recognition-of-active-inflammatory-changes-compatible-with-axial-spondyloarthritis-on-magnetic-resonance-imaging-of-sacroiliac-joints/