Background: Effusion-(ES) and Hoffa-synovitis (HS) are established features for evaluating structural pathology by the semiquantitative (sq) MOAKS system. ES severity, and the quantitative (q) MRI signal and texture characteristics (radiomics) of the Hoffa’s fat pad (HFP) are often used as sensitive, albeit not highly specific, surrogates of whole-joint inflammation, in the absence of contrast enhancement (CE).

Objectives: To develop a fully automated pipeline for ES volume morphometry (qESV) and HS radiomics. Further, to assessed their utility for predicting (combined symptomatic and radiographic) osteoarthritis (OA) progression, compared with expert sq MOAKS ES/HS readings.

Methods: Automated segmentation of peripatellar ES and the HFP was performed by convolutional neural networks (CNN), trained on >100 manual segmentations. The qESV, and about 100 established radiomic signal and texture parameters in various HFP regions of interest (ROIs) were evaluated from sagittal IW TSE FS MRIs of the knee in the FNIH Biomarker Study of the OA Initiative: 194 displayed combined radiographic and pain progression over 2-4 years; 200 did not show either type of progression. Cohen’s d ( d ) served as a measure of effect size of differences between progressors and non-progressors. Radiomic measures with the highest d values were submitted to binary logistic regression (odds ratios normalized to the control SD [qORs]), adjusting for covariates and with “progression status” as dependent variable. For validation purposes, q measures were converted into ordinal (sq) CNN-ES and HS scores, and the resulting sqORs for progressor status compared with expert sq ES and HS MOAKS readings. We assumed a monotonic ordering of categories, and the sqORs to represent a one-category increase.

Results: The qESV displayed a d of 0.24 (progressors vs. non-progressors). Combining all HFP radiomic measures, the d tended to be higher in the posterior 10% of the HFP adjacent to the synovial membrane, but similar for all medial-lateral ROIs (0.21 for all, 0.20 for patellar, and 0.19 for 5 central slices). The d for the entire Hoffa was slightly lower (0.18, 0.17, 0.16).

The qESV displayed a qOR of 1.24 (95% CI 1.01, 1.51) and the sqORs for ES were 1.45 (95% CI 1.08, 1.94) for CNN-derived vs. and 1.11 (95% CI 0.86, 1.43) for expert MOAKS sq scores. The relatively greatest radiomic qOR (“Kurtosis of MRI signal” in the posterior 10%ROI/ patellar slices only) was 1.77 (95% CI 1.37, 2.27). The corresponding sqORs for HS were 2.38 (95% CI 1.63, 3.49) for CNN-derived vs. 1.96 (95% CI 1.38, 2.77) for expert MOAKS sq scores. Most progressor/non-progressor comparisons were statistically significant, with 95% CIs not crossing 1.0.

Conclusions: We here evaluate the prognostic performance of ES morphometry and HFP radiomics with respect to OA progression status. The study relied on fully automated segmentation of ES and the HFP using CNNs. A posterior HFP ROI was observed to provide the relatively strongest difference of radiomic measures between progressors and non-progressors. This is likely due to a closer proximity to the synovial membrane, with signal and texture measures potentially more specific to HS than anteriorly. qESV morphometry and HS radiomics revealed that, after conversion to ordinal scores, sqORs were at least as high as those derived from sq expert MOAKS reading, predicting OA progression. The results suggest that a CNN-based automated pipeline can provide scalable, q biomarkers of inflammation (ES and HS) as potential surrogate endpoints in clinical studies, without necessarily employing CE-MRI.

REFERENCES: NIL.

Acknowledgments: NIL.

Disclosure of Interests: Felix Eckstein Chondrometrics GmbH, Chondrometrics GmbH, Kolon TissueGene (KTG), Galapagos, Novartis, 4P Pharma/4 Motion, Trialspark/Formation Bio, Peptinov, Sanofi, Artialis, Argenx., Kolon TissueGene (KTG), Galapagos, Novartis, Peptinov, Sanofi, Kowsar Sheikhi Valashani Chondrometrics GmbH, Tobias Winkler Pluri Biotech Ltd., PROTO Coordinator, Jamie Collins Chondrometrics GmbH, Frank Roemer Boston Imaging Core Lab, Boston Imaging Core Lab, Ali Guermazi Boston Imaging Core Lab, Boston Imaging Core Lab, Kolon TissueGene (KTG), Novartis, 4P Pharma, Formation Bio, Peptinov, Kolon TissueGene (KTG), Novartis, David Hunter Kolon Tissue Gene, Novartis, Kolon Tissue Gene, Novartis, Wolfgang Wirth Chondrometrics GmbH, Chondrometrics GmbH.