IDENTIFICATION OF MACHINE-INFORMED REGIONAL HOTSPOTS WITHIN MAGNETIC RESONANCE IMAGES ASSOCIATED WITH KNEE...

It is difficult to identify the origin of knee OA as it is a heterogeneous disease, with the subchondral bone being a relatively neglected tissue. The current bottom-up hypothesis-driven process is costly, slow, and sometimes results in failure of a biomarker that scientists have spent many years refining. Therefore, employing a top-down approach to randomly select as many regions of interests and associated imaging features as possible to converge on potential hotspots exhibiting multiple clinical correlations could enhance our ability to pinpoint high priority symptom-correlated regions for future hypothesis-testing poised for success.

Discover subchondral volumes of interest (VOI) and radiomic knee MRI features that represent underlying physiological causes of knee OA symptoms by applying high-dimensional methods.

This study used MRIs and clinical data of 2804 women and 1992 men (n=4796), 45-79 years old from the OAI. Images were acquired in double-bullseye orientation using coronal IM FSE MRI at 3T, 0mm gap, with voxel size of 0.365 × 0.456 × 3.0 mm³. A 3D anatomical MRI atlas of the knee bone structure was generated by combining MRI scans for each KLG (0-4), and sex (male and female) until the mean absolute change in voxel difference reached a plateau. Five watershed configurations (maximum settings of 10, 100, 250, 500, and 750 VOIs per bone) were applied to the atlas to define a large array of contrast-guided VOIs within the femur and tibia. The watershed-labeled atlas was co-registered to each participant's knee MRI and radiomic features for each VOI was extracted using PyRadiomics. Multilinear principal component analysis was used to reduce dimensionality of the 107 2D matrix-structured radiomics data, yielding latent factor scores, an approach previously applied with brain functional MRI. Radiomic latent factor scores were related to Knee Osteoarthritis Outcome Score (KOOS) subscales measured at 10 timepoints across 10 years. Associations were modelled in a random-effects linear regression for knee-level subscales, accounting for age, sex, body mass index, KLG, physical activity, income, education level, alcohol use, smoking, comorbidities, depression, hormone replacement, and glucocorticoid use. Standardized estimates per watershed VOI, and per set of radiomic latent factor scores were mapped back to the knee atlas and overlaid to compute mean absolute effect estimates representing parametric maps per timepoint. Hotspots were defined as regions with the highest average absolute effect, indicating significant associations with multiple outcomes.

For the KOOS symptoms and pain subscales, the timepoint with the largest overall mean effect size was at baseline. Hotspots for both subscales were localized to the mid-anterior aspect of the medial tibial condyle just inferior to the subchondral plate, with average absolute effect estimates across all radiomic features and watershed configurations of 0.91 [0.86, 1.06] for symptoms and 1.11 [1.01, 1.48] for pain. The radiomic feature that contributed the largest effect within the hotspot was neighboring gray tone difference matrix (NGTDM) strength, a measure of spatial rate of intensity change. On average, the absolute magnitude of association between local distinct gray-level voxel groups in this region was 2.15 [2.22, 2.49] and 2.51 [2.65, 3.39] unit difference for KOOS symptoms and pain, respectively.

Multiple MRI-derived subchondral bone radiomic features showed the strongest associations with knee OA symptoms and pain within the mid-anterior aspect of the medial tibial condyle just inferior to the subchondral plate, suggesting this region is a high priority potential contributor to early knee OA outcomes. Prominent local gray-level patterns may represent a key radiomic marker of symptoms and pain burden.