AUTOMATIC SEGMENTATION-FREE INFLAMMATION SCORING IN RHEUMATOID ARTHRITIS FROM MRI USING DEEP LEARNING

Background: Visual scoring using RAMRIS is commonly used to assess inflammation within extremity MRI scans from patients with (suspected) rheumatoid arthritis (RA). As visual scoring is time-consuming and requires rigorous training, automatic scoring would accelerate and facilitate RA diagnosis and monitoring. While image interpretation by deep learning (DL) might provide efficient and accurate scoring, previous DL methods rely on accurate segmentations of all relevant anatomical structures, which is also a time-consuming process, prone to propagation of errors. Objectives: A DL method is needed to automatically analyze extremity MRI scans to score inflammatory signs, without the need for accurate segmentation. Methods: MRI scans of the hands and feet of 2148 subjects (727 patients with clinically suspect arthralgia (CSA), 1247 patients with early onset arthritis and 174 healthy controls) were collected for developing the method (394 as a held-out test set, the remaining for 5-fold cross-validation). Scans of 127 CSA patients from the TREAT-EARLIER trial were used for independent validation. Both datasets include signs of (teno-)synovitis and bone marrow edema (BME). A DL model was developed to perform RAMRIS scoring, focusing on inflammation scores (see Figure 1a). Model performance was assessed using Pearson’s correlation coefficient and intra-class correlation coefficient (ICC) as compared to the mean inflammation scores from two readers. A heatmap was developed to interpret how the model operates, by indicating the most informative areas. Results: Table 1 shows the correlations and ICCs of the proposed method on the test- and validation-set. The model’s output correlated strongly with total visual scoring of synovitis and tenosynovitis with R=0.86/0.83 and ICC=0.75/0.64. For BME, these metrics were 0.65 and 0.47, respectively. The performance on the independent validation set decreased by 0.1 to 0.2 for synovitis and tenosynovitis and over 0.4 for BME. The heatmaps showed that the DL model’s output was mainly -but not exclusively- based on the specific RAMRIS areas (Figure 1c). Conclusion: Fully-automatic inflammation scoring by DL interpretation of MRI scans is feasible, especially in scoring (teno-) synovitis, which is one of the most important features of early RA, in the wrist and MCP. The DL model could be developed much quicker than methods that require accurate segmentations prior to quantification. Since the heatmaps highlighted not only inflammatory areas corresponding to the RAMRIS definition, but also some other areas, the proposed DL-based inflammation scoring method should be validated further and improved in the future. Table 1. The mean Pearson’s correlation coefficients (±SD) and ICC (±SD) on the test set and independent validation set. Most results were statistically significant (p value < 0.05), except for coefficients less than 0.7. In bold: Corr./ICC over 0.6, in italic: Corr./ICC less than 0.4. Dataset Score (from wrist, MCP and MTP) Corr. (SD) ICC (SD) Test Total inflammation 0.88 (±0.03) 0.79 (±0.03) (rough reference, 2 observers) 0.86 0.85 Synovitis 0.86 (±0.05) 0.75 (±0.07) Tenosynovitis 0.83 (±0.06) 0.64 (±0.12) BME 0.65 (±0.11) 0.47 (±0.05) Validation Total inflammation 0.73 (±0.03) 0.36 (±0.06) (rough reference, 2 observers) 0.83 0.91 Synovitis 0.74 (±0.02) 0.55 (±0.04) Tenosynovitis 0.61 (±0.06) 0.57 (±0.04) BME 0.12 (±0.04) 0.05 (±0.02) Figure 1. a) overall workflow of the DL model; b) scatter plots of DL vs visual scoring for example in the wrist; green line: identity; red line: regression; and c) examples of heatmaps for tenosynovitis in the wrist. REFERENCES: NIL. Acknowledgements: This AI-project has been funded by the Dutch Research Council (NWO) Applied and Engineering Sciences (project number 17970). The TREAT-EARLIER trial has been funded by an NWO-ZonMW grant (project number 95104004). The Dutch Arthritis Society contributed financially to both grants. Disclosure of Interests: Yanli Li: None declared, 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., 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., Dennis A. Ton: 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 H.M. 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. DOI: 10.1136/annrheumdis-2024-eular.2951 Keywords: Magnetic Resonance Imaging, Outcome measures, Artificial Intelligence, Imaging Citation: , volume 83, supplement 1, year 2024, page 730Session: Rheumatoid arthritis (Poster View)

Magnetic Resonance Imaging, Outcome measures, Artificial Intelligence, Imaging