DYNAMIC CONTRAST-ENHANCED MRI OF THE SYNOVIUM IN KNEE OSTEOARTHRITIS: SEMI-AUTOMATIC SEGMENTATION OF SYNOVIAL SUBREGIONS AND TEST-RETEST REPEATABILITY

Abstract

Inflammation of the joint lining (synovitis) is often seen in OA, and is strongly associated with both knee pain severity and disease progression. Dynamic contrast-enhanced MRI (DCE-MRI) can be used to visualize the uptake and washout of gadolinium-based contrast, and enables quantification of blood perfusion as a surrogate measure of inflammation. DCE-MRI could therefore be used to characterize the extent of synovitis and serve as a biomarker for response to treatment in clinical trials. However, (semi)-automatic segmentation methods and thorough technical and clinical validation are needed to fully adopt DCE-MRI in this setting. We aim to evaluate semi-automatic segmentation of synovial subregions and test-retest repeatability of DCE-MRI to measure synovial perfusion in patients with knee OA. Patients with mild-to-moderate knee OA (KLG 1-3) who were allocated to the control group in a prospective randomized controlled trial of genicular artery embolization were included. Patients underwent MRI at 3T (GE Healthcare) at baseline and after 1 month. Image acquisition consisted of native T1 mapping with variable flip angle, followed by a dual-echo SPGR sequence (DISCO; DIfferential Sub-sampling with Cartesian Ordering), in which the spatial resolution was optimized to allow for visualization of small vessels. After the initial phase, 0.1 ml/kg gadovist was injected and subsequently 34 dynamic phases were acquired, with a temporal resolution of 10.2 seconds. Segmentation of the synovium comprised an initial rough manual segmentation followed by selection of enhancing voxels using a shuffle transform method. Synovial subregions were created by semi-automatic vessel mapping using MeVisLab and mathematically assigning synovial voxels to the corresponding closest genicular artery. Motion compensation was performed using Elastix (Elastix, version 5.0.1), and native T1 mapping and pharmacokinetic modelling was done using open source software for DCE analysis (MADYM, version 4.21.1). The commonly used combination of the Parker literature arterial input function, and the Extended Toft's model pharmacokinetic model was implemented. Voxelwise perfusion parameters were calculated, after which median Ktrans values were extracted for the whole synovium and separately for each synovial subregion. Agreement was visualized with a Bland-Altman plot, and test-retest repeatability was evaluated using intraclass correlation coefficients (ICC; single rating random sample) and within-subject coefficients of variation (CV). 30 patients were included, of which two patients were lost to follow-up and one patient was excluded due to errors during image acquisition. For the remaining 27 participants, median Ktrans values ranged from 0.017 to 0.121. ICC of Ktrans for the whole synovium was 0.51, and CV was 0.14. Sensitivity analysis by exclusion of two statistical outliers yielded an ICC of 0.66 and a CV of 0.11. For the individual synovial regions, ICCs ranged between 0.27 and 0.58 while CVs ranged between 0.13 and 0.25. Semi-automatic segmentation of the synovium and synovial subregions through vessel mapping on DCE-MRI in knee OA is feasible. Perfusion parameters determined through pharmacokinetic modeling have moderate repeatability and should be interpreted with caution. Stichting Coolsingel, Cook Medical, Erasmus MRace Doelmatigheid E.H.G. Oei has received research support from GE Healthcare. The University of Wisconsin receives research support from Bracco Diagnostics. CORRESPONDENCE ADDRESS: [email protected]