MRI-based radiomics to determine the role of thigh muscle quality in the development of knee osteoarthritis: A discovery and validation study.

introductionAlthough the reduced muscle quality is common in elderly people, its role in the development of knee osteoarthritis (OA) remains unclear. Magnetic Resonance Imaging (MRI) is considered the gold standard imaging modality to assess muscle quality, which might holistically capture the underlying pathology.

objectivesThis study aimed to apply MRI-based radiomics to assess the quality of thigh muscles comprehensively, and examine its prediction performance for incident knee OA.

methodsThe participants at risk of knee OA during the 48-month visit were selected from a multicenter cohort and analyzed using a nested case-control design. Cases were matched 1:1 to controls according to age, sex, and contralateral knee status. A multitask framework was developed to comprehensively evaluate thigh muscle quality by calculating conventional quantitative markers and radiomic features. Baseline collected thigh and knee MRI scans were respectively used for discovery and validation procedures. In each procedure, three (cross-sectional area [CSA], intramuscular adipose tissue [intra-MAT] CSA, and radiomic signature score [Rad-score]) imaging models of each muscle group (extensors, flexors, and auxiliaries) were trained and tested. The prediction performance of each muscle Rad-score was compared with conventional markers respectively. Furthermore, the incremental predictive value of clinical characteristics for Rad-score model was investigated.

results448 participants (mean age 61 years, 63 % female) were identified and split into training (314 [70.1 %]) and testing (134 [29.9 %]) cohorts. In the discovery procedure, the extensors Rad-score provided superior prediction performance than extensors CSA and intra-MAT CSA (area under the curve [AUC] 0.832 [95 % CI 0.765-0.899] vs 0.595 [0.498-0.692] and 0.600 [0.503-0.697] in testing cohort; all p < 0.001). Contrarily, the performances of the Rad-score and conventional markers of flexors were relatively poor, as were those of the auxiliaries. Extensors Rad-score also outperformed auxiliaries Rad-score, flexors Rad-score, and the clinical model (AUC 0.832 [0.765-0.899] vs 0.687 [0.597-0.777], 0.634 [0.541-0.728], and 0.577 [0.483-0.671] in testing cohort; all p < 0.001). Unexpectedly, the integration of other radiomic signatures and clinical characteristics showed comparable net benefit in decision curve analysis when compared with extensors Rad-score alone. Similar findings were obtained in the validation procedure.

conclusionMRI-based radiomics could enable comprehensive assessment of muscle quality and allow exact prediction of incident knee OA, which helps uncover the precursory role of reduced extensors quality in developing knee OA. The radiomic signature from extensors would be an alternative indicator for future diagnosis and treatment in knee OA.