Background: On non-contrast enhanced MRI, infrapatellar or “Hoffa”-synovitis (HS) may serve as a sensitive (albeit non-specific surrogate) of whole joint inflammation in osteoarthritis (OA). The current gold standard is semiquantitative (sq) radiological expert reading based on the MOAKS system. To date, it is unclear whether scalable, fully automated, quantitative (q) imaging markers of synovitis are clinically useful.

Objectives: To elucidate whether q imaging measures of HS predict a relevant and “hard” (objective) clinical outcome, i.e. knee arthroplasty (KA). Specifically, we evaluated the prognostic value of complex q MRI signal and texture measures (radiomics) of Hoffa’s fat pad (HFP) for subsequent KA, compared with sq HS MOAKS scores.

Methods: Fully automated HFP segmentation was performed on sagittal IW TSE FS MRIs, using a convolutional neural network (CNN) trained on 160 manual HFP segmentations, in a different cohort. Approximately 100 q signal intensity and texture radiomic parameters, recommended by the Image Biomarker Standardization Initiative (IBSI), were compared between regions of interest (ROIs): (i) all sagittal HFP slices, (ii) HFP slices displaying the patella, and (iii) five central patellar slices. ROIs encompassed percent regions of depth (posterior to anterior). Over 5 years, 199 knees (of 4796 OA Initiative participants) underwent KA. These were matched with 199 controls, based on radiographic stage, sex, age, and time point of KA, who did not receive KA. MRIs acquired at the last time-point before KA (T0) and two years earlier (T _-2 ) were studied. Conditional logistic regression was used to assess the prognostic value of the q radiomic and sq HS assessment (unadjusted ccOR). For validation purposes, the q measures were converted into sq ordinal CNN-HS scores and were compared with expert MOAKS HS reading.

Results: Participants were 64±8.4 years old and 58% women. At T ₀ , the greatest difference between KA cases and controls, across radiomic measures and HFP ROIs, was 0.71 (Cohen’s D), and 2.01 (q ccOR; 95%CI 1.64, 2.46). The performance across measures and ROIs was heterogeneous. The CNN-HS score revealed a sq ccOR of 2.48 (1.90, 3.24), and the expert MOAKS HS grades one of 2.16 (1.65, 2.83). At T _-2 , the Cohen’s Ds were observed to be lower (<0.50). The sq ccOR for CNN-HS was 1.54 (1.19, 2.29) and that for expert MOAKS HS scores was 1.64 (1.17, 2.29). These differences were statistically significant at both T ₀ and T _-2.

Conclusions: We have developed a fully automated pipeline for extracting complex and q measures of HS, using non-contrast enhanced MRI. Further, we assessed the prognostic performance of these radiomic features in relation to surgical KA. We find these HS endpoints to significantly differentiate between KA cases and controls. This occurs with greater accuracy at the time point before KA than two years prior. No single ROI appeared to predict KA status superiorly to others. Comparison of CNN-based sq HS with expert MOAKS HS scores was carried out (only) for validation purposes, as it permits applying the same statistical measure (sq ccOR), and revealed similar performance for both approaches. This suggests that complex q parameters of HFP MRI signal and texture (radiomics), derived automatically from an AI (CNN-based) algorithm, can provide a scalable imaging biomarker of inflammation, and potential surrogate endpoints in clinical studies of OA. Further work should identify to what extent the combination with different q measures of other tissues can further improve KA prediction.

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, Tobias Winkler Pluri Biotech Ltd., PROTO Coordinator, C Kent Kwoh: None declared, Frank Roemer Boston Imaging Core Lab LLC, Boston Imaging Core Lab LLC, Ali Guermazi Boston Imaging Core Lab (LLC), Boston Imaging Core Lab (LLC), Kolon TissueGene (KTG), Novartis, 4P Pharma, Formation Bio, Peptinov, Kolon TissueGene (KTG), Novartis, David Hunter Kolon TissueGene (KTG), Novartis, Wolfgang Wirth Chondrometrics GmbH, Chondrometrics GmbH.