STATISTICAL SHAPE MODELING OF THE HIP JOINT SPACE CAPTURES RADIOGRAPHIC SEVERITY AND PROGRESSION PATTERNS

The definition of radiographic hip OA (RHOA) is still dependent on semi-quantitative grading systems such as the KLG and the (modified) Croft. These semi‑quantitative systems rely heavily on reader expertise, limiting reproducibility and comparability across studies. JSW is currently the most reliable radiographic feature for describing hip OA, and continuous JSW measurements offer a promising route toward automated, objective, and reproducible RHOA assessment. Joint space shape and change over time could provide an opportunity to quantify structural variation and progression in a continuous manner without reducing the joint space to a single measure. However, it remains unclear whether specific joint space shape can capture longitudinal change, OA severity, or both.

To investigate whether joint space shape derived from longitudinal anteroposterior (AP) pelvic radiographs can capture 1) time‑dependent changes, 2) RHOA severity, or 3) severity‑dependent progression patterns.

We included 955 participants (6754 observations) from the prospective Cohort Hip and Cohort Knee (CHECK) study, a longitudinal cohort with follow-up at 2, 5, 8 and 10 years, in our exploratory study. The mean baseline age of the participants was 55 ± 5.2 years and 79% was female. The joint was automatically segmented from the pelvic radiographs using a pre-trained U-Net architecture, which was fine-tuned for hip segmentation. From the segmentation mask, the joint space was described using 100 landmarks, see Figure 1a. These landmarks were used to create a statistical shape model (SSM) to describe joint space shape variation in our sample. Shape modes explaining ≥ 2% of total variation were selected for analysis. RHOA was quantified using KLG. The KLG was categorized into an ordinal 3-level scale: "free of OA" (KLG = 0), "doubtful OA" (KLG = 1), and "definite OA" (KLG ≥ 2). Recategorization of the KLG was done to create a balanced distribution, with 38.5% of observations free of OA, 35.2% doubtful OA, and 26.3% definite OA. Hips with total hip replacements where excluded. Linear mixed-effects models were fitted for each shape mode to investigate 1) time dependency of the shape mode, 2) relationship with RHOA severity, 3) severity‑dependent progression patterns modeled with and without an interaction between time and RHOA. Mixed effects were included to account for hip side and individual. Model comparisons were performed using likelihood ratio tests. Polynomial contrasts (linear/quadratic) were used to model ordinal RHOA.

The first two shape modes explained 98% of total shape variance in the population. A visualization of the shape modes 1 and 2 is shown in Figure 1b and 1c, respectively. Shape mode 1 changed significantly over time (p < 0.001), as well as with RHOA severity (linear: p < 0.001). The association with severity disappeared after adjusting for time and no interaction with severity was present. The model including only time showed the best fit for shape mode 1. Shape mode 2 was associated with RHOA severity (linear: p < 0.001) and also changed over time (p < 0.001). The model including time, RHOA and their interaction showed the best fit for shape mode 2. Predicted shape mode 2 trajectories by RHOA severity are shown in Figure 2.

This exploratory study quantified two distinct structural phenotypes of joint shape using SSM. Shape mode 1 reflects a longitudinal progression pattern independent of RHOA severity, however, it could also partly reflect other patterns not explored in this study. Shape mode 2 seems to captures both RHOA severity and severity-dependent progression trajectories. These findings highlight the potential of joint space-based shape modeling for automated, objective RHOA grading and for identifying structural phenotypes relevant to disease progression.