PROGNOSTIC VALIDITY AND SENSITIVITY TO CHANGE OF AUTOMATICALLY MEASURED HOFFA RADIOMICS IN PROGRESSOR VS....

High signal intensity alterations on non-enhanced fluid-sensitive fat suppressed MRI can be observed in the intracapsular and extrasynovial Hoffa's fat pad. These are commonly used as a proxy for whole-joint inflammation in studies lacking direct synovitis visualization using contrast-enhancement (CE). Radiomics extract mathematical constructs from imaging data that are not accessible to the human eye; these have been shown to increase the prediction of predefined outcomes, particularly in the field of oncology.

To evaluate the cross-sectional and longitudinal prognostic validity of Hoffa radiomics for predicting OA progression and their sensitivity to longitudinal change over two years.

We analyzed the OAI FNIH-1 data: 194 knees with combined radiographic and pain progression (ComP), 103 with radiographic-only (ROP), 103 with pain-only (POP), and 200 without progression (NonP). Hoffa's fat pad was automatically segmented using sagittal IW TSE MRIs of all 600 knees at baseline (BL), year 1 (Y1), and year 2 (Y2), using a U-Net trained on 160 manually segmented knees. Seventy-five texture and 19 signal intensity measures were extracted across all MRI slices covering Hoffa's fat pad, focusing on its posterior 10%, i.e., the region adjacent to the synovial membrane (Fig. 1). A multi-component predictor, termed "quantitative progression probability" (QPP), was developed using stepwise logistic regression, to distinguish ComP from NonP. This was based on measures obtained at the intermediate Y1 visit, allowing unbiased application to BL, Y2, and to BL→Y2 change. The most sensitive texture and signal intensity features were identified across BL, Y2, and BL→Y2 based on Cohen's d between ComP and NonP. Longitudinal change and standardized response mean (SRM) of the most sensitive texture and signal intensity measures between BL and Y2 were assessed. Odds ratios (ORs, normalized to the SD) for ComP, ROP, and POP (each vs. NonP) were estimated for these measures and for QPP using logistic regression, adjusted for baseline KLG, minimum JSW, pain medication use, WOMAC pain, age, sex, BMI, and race.

The QPP was based on two gray-level dependence matrix (GLDM) measures (GLDM variance and dependence entropy). The measures most sensitive to differences between ComP and NonP were the interquartile range of the signal intensity (SI-IQR) and the zone entropy of the gray-level size zone matrix (ZE-GLSZM). All three measures predicted ComP vs. NonP similarly at both BL and Y2 (ORs 1.5 to 2.2, Table 1). QPP and SI-IQR also predicted ROP vs. NonP at both visits (ORs 1.3 to 1.8), whereas ZE-GLSZM predicted ROP vs. NonP only at Y2 (OR 1.7, Table 1). Hoffa radiomic measures were not prognostic of POP vs. NonP (Table 1). ORs were somewhat smaller in longitudinal than in cross-sectional analyses, with the strongest ORs for BL→Y2 change observed for ZE-GLSZM (OR 1.4, Table 1). At both cross-sectional visits, SI-IQR (Fig. 2) and ZE-GLSZM (data not shown) values were generally greater in ComP and ROP than in POP and NonP knees. The SRM tended to be greater for SI-IQR than for ZE-GLSZM with the largest SRM observed for 2-year change in SI-IQR in ROP knees (SRM=0.96).

Hoffa MRI radiomic measures appear to be associated with combined and isolated radiographic progression, but not isolated pain progression in knee OA. This supports their role as quantitative prognostic markers of structural disease activity in the absence of CE MRI. Their high responsiveness (sensitivity to change) suggests utility for monitoring disease progression in clinical trials, pending validation of their responsiveness to treatment (predictive validity).