PSSF: Early osteoarthritis detection using physical synthetic knee X-ray scans and AI radiomics models

<h2>Abstract</h2><h3>Purpose</h3> To develop a physics-based synthetic simulation framework (PSSF) for generating controllable anteroposterior knee radiographs and to use it as a methodological testbed for radiomics feature stability and machine-learning robustness under acquisition domain shifts. <h3>Materials and methods</h3> A 2D X-ray projection simulator was built from a parametric knee anatomy model with KL-like grade mapping (0–2) via joint space width, osteophytes, sclerosis proxy, and alignment parameters. A virtual cohort of 180 subjects (260 knees; 780 radiographs) was generated and imaged under three protocols (reference, low-dose, and geometry-shift). A 512×512 medial joint region was automatically localized; 107 radiomic features were then computed with an IBSI-aligned implementation following standardized definitions and preprocessing. Logistic regression, random forest, and gradient boosting models were trained for binary (KL-like "0" vs. "2") and three-class (0–2) classification under within-protocol, cross-protocol, and multi-protocol training scenarios. Feature stability was quantified using intraclass correlation coefficients (ICC[2,1]). <h3>Results</h3> For the primary binary task, within-protocol training/testing achieved AUC=0.92 and balanced accuracy=0.92. Cross-protocol testing (reference-trained models evaluated on low-dose and geometry-shift images) reduced performance to AUC=0.78 (balanced accuracy=0.78). Multi-protocol training improved generalization across acquisition conditions (AUC ≈ 0.88–0.90; balanced accuracy ≈ 0.88–0.90 across protocols). Across all extracted radiomic features, 43/107 (40.2%) demonstrated high stability (ICC ≥ 0.90) across acquisitions, forming a candidate robust feature subset. <h3>Conclusion</h3> Physics-based, parameter-controllable synthetic radiographs enable systematic stress-testing of radiomics pipelines and ML models under known acquisition perturbations. This study is strictly methodological and confined to a synthetic domain; clinical translation requires calibration and external validation on real-world datasets.