A DEEP LEARNING MODEL TO LOCATE INFLAMMATORY CHANGES IN RHEUMATOID ARTHRITIS

T. Hassanzadeh, D. Shamonin, Y. LI, M. Reijnierse, A. Van der Helm – van Mil, B. StoelLeiden University Medical Center, Division of Image Processing, Department of Radiology, Leiden, Netherlands Leiden University Medical Center, Department of Radiology, Leiden, Netherlands Leiden University Medical Center, Department of Rheumatology, Leiden, Netherlands  Background Magnetic Resonance Imaging (MRI) is one of the most sensitive imaging modalities to monitor affected joints and detect inflammatory signs related to Rheumatoid Arthritis (RA). Finding and tracking inflamed areas can help to evaluate patients status, and start treatment in a very early stage. The RAMRIS [1] scoring system is being used to assess the extent of inflammation by scoring specific areas to quantify synovitis, tenosynovitis and bone marrow edema. Deep Learning (DL) may, however, provide more accurate and specific quantifications. Objectives We aimed to develop a change map that locates RA progression over time, using MRI scans of the wrist. Our purpose was to develop an automatic, hypothesis-free DL model that considers the entire MRI for detecting all possible longitudinal changes related to RA between baseline and follow-up wrist MRIs. Methods Wrist contrast-enhanced T1-weighted fat-suppressed MRI scans of 236 patients from four time points (baseline, with 4, 12, and 24 months follow-up) were collected [1]. An unsupervised DL model was developed together with image processing techniques to detect a pixel-level progression within wrist MRI scans between baseline and the last follow-up MRIs. This model predicts the MRI appearance of the follow-up scan, based on the baseline scan, by learning from MRI time sequences of all patients in the training set. Similarly, the baseline scan was predicted backwards in time, based on the follow-up scan (see Figure 1). Subsequently, by comparing the predicted and true MRI image, pixel-level changes can be determined between the two MRI scans. For prediction, we used a U-Net [2] to reconstruct follow-up images from baseline and vice versa. After pre-training, two copies of the trained model and a joint loss function were used for final training (a dual branch U-Net); the first branch is trained to generate a follow-up image and the second branch to create a baseline image. After final training, two difference maps were obtained by subtracting the baseline image from the predicted follow-up image and by subtracting the follow-up image from the predicted baseline image. Since real changes appear in pixels with positive values, Otsu thresholding was applied only on the positive differences to generate two change maps, where the first change map shows intensity reduction (areas of remission) and the second indicates increased intensity (areas of progression). With fusion of the obtained change maps we obtained the final change map that can highlight the changes in both directions. Results As can be seen from Figure 1, the proposed model could find local pixel-level changes, highlighting RA remission/progression over time. The red overlay shows a reduction in intensity, the blue overlay indicate an increase in intensity. To show the performance of the proposed model, the results were compared with a baseline model, that simply subtracts the original follow-up from the baseline image. As can be seen from the provided examples, the proposed model is less sensitive to imaging artifacts. Conclusion A comparative DL-model is proposed to find inflammatory changes in wrist MRI scans. Despite being an unsupervised model (sometimes detecting changes, in e.g. vessels, that are presumed to be unrelated to RA), the results are promising, since the change maps contain less artifacts as compared to simple image subtraction. References M. Østergaard, et al., “Omeract rheumatoid arthritis magnetic resonance imaging studies. core set of mri acquisitions, joint pathology definitions, and the omeract ra-mri scoring system.” The Journal of rheumatology, vol. 30, no. 6, pp. 1385–1386, 2003. O. Ronneberger, et al., “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241. Image/graph:Figure 1. Overview of the proposed model and examples of obtained change maps. Acknowledgements This AI-project has been funded by the Dutch Research Council (NWO) Applied and Engineering Sciences (project number 17970). The TREAT-EALRIER trial has been funded by an NWO-ZonMW grant (project number 95104004). The Dutch Arthritis Society contributed financially to both grants. Disclosure of Interests Tahereh Hassanzadeh Grant/research support from: Bristol-Myers Squibb and Pfizer contributed to this project, through a grant from the Dutch Research Council (NWO), Applied and Engineering Sciences., Denis Shamonin Grant/research support from: Bristol-Myers Squibb and Pfizer contributed to this project, through a grant from the Dutch Research Council (NWO), Applied and Engineering Sciences., Yanli Li: None declared, Monique Reijnierse Grant/research support from: Bristol-Myers Squibb and Pfizer contributed to this project, through a grant from the Dutch Research Council (NWO), Applied and Engineering Sciences., Annette van der Helm – van Mil Grant/research support from: Bristol-Myers Squibb and Pfizer contributed to this project, through a grant from the Dutch Research Council (NWO), Applied and Engineering Sciences., Berend Stoel Grant/research support from: Bristol-Myers Squibb and Pfizer contributed to this project, through a grant from the Dutch Research Council (NWO), Applied and Engineering Sciences. Keywords: Rheumatoid arthritis, Artificial Intelligence, Imaging DOI: 10.1136/annrheumdis-2023-eular.500Citation: , volume 82, supplement 1, year 2023, page 298Session: New developments in imaging in rheumatology (Poster Tours)