POS0900 AUTOMATIC SCORING OF EROSION, SYNOVITIS AND BONE OEDEMA IN RHEUMATOID ARTHRITIS USING DEEP LEARNING ON HAND MAGNETIC RESONANCE IMAGING

Background Rheumatoid Arthritis (RA) Magnetic Resonance Imaging (MRI) scoring system (RAMRIS) [1] is used to manually assess severity of disease activity and monitor treatment response, but it dependent on observer variability time-consuming. Deep learning techniques have the potential improve reproducibility efficiency RAMRIS by automating analysis hand MRI scans, however, there are limited data an automated assessment approach. Objectives To investigate whether a deep neural network (DNN) can be trained automatically detect erosion, synovitis, oedema in RA patients using scans RAMRIS. Methods We 1,5 Tesla (Siemens Magnetom Vida Aera) from BARE BONE trial, prospective, single-arm, interventional, open-label, phase 4 trial (EUDRACT 2018-001164-32) which were treated with baricitinib (4 mg/day) for 48 weeks. One objectives was effect joint damage synovial inflammation. Participants received at week 0, 24, following standardized scanning protocol [2]. All images scored according DNNs applied coronal T1 (pre/post contrast enhancement) T2 MR images. 3-D landmarks each location identified region interest (ROI) around landmark extracted train DNN. Three separate trained, one subcomponents (erosion, oedema). Each DNN based ResNet-3D [3] architecture that pretrained video classification task [4]. The networks predict scores characteristic into three classes ranging 0 (no pathological change) 2 (high burden). performance evaluated area under receiver operating curve (AUROC) precision-recall (PR-AUC). Three-fold cross-validation hold-out test set evaluated. Results In total, we obtained 212 both weighting 30 (24 woman/6 men, age 53.5±12.6 years, duration 4.3±4.4 years). overall score decreased 20.6 (CI95% 14.4 27.8) 18.3 11.5 26.5) 48. For evaluation erosions oedema, 23 respectively per hand, 7 synovitis. total 4608 erosion available, synovitis 1152 landmarks. AUROC predicting 86±2% PR-AUC 83±4%. prediction 78±14% 83%±10%. Despite low number ROI scoring, respective 60±4% 69±3%. Conclusion This proof-of-concept study demonstrated fully extraction bone feasible. future, our approach may help routine clinical practice trials high accuracy while keeping costs human resources manageable. References [1]Østergaard M et al. Journal Rheumatology 2017. [2]Kemenes S Annals Rheumatic Diseases 2022; 81:1320-1321. [3]Hara K arXiv:1711.09577, 2018. [4]Kay W https://arxiv.org/pdf/1705.06950 . Figure 1. Neural pipeline. Region interests required fed ResNet 3D respectively. Acknowledgements supported Lilly Deutschland GmbH DFG (FOR2886 PANDORA, CRC1181, CRC1483 EmpkinS). Additional funding BMBF (project MASCARA), ERC Synergy grant 4D Nanoscope, IMI funded project RTCure, Emerging Fields Initiative MIRACLE FAU Erlangen-Nürnberg. M.S., J.Q. K.B. gratefully acknowledge support d.hip campus - Bavarian aim. Infrastructural hardware provided Digital Health Innovation Platform. Disclosure Interests Maja Schlereth: None declared, Arnd Kleyer Consultant of: Received consulting fees GmbH., Jonas Utz: Lukas Folle: Sara Bayat Filippo Fagni: Ioanna Minopoulou: Koray Tascilar Jule Taubmann: Michael Uder: Tobias Heimann: Jingna Qiu: Georg Schett: Katharina Breininger: David Simon GmbH.