Abstract Background Stereotactic body radiation therapy (SBRT) delivers ablative doses with high precision but is sensitive to target motion, especially rotational errors. Conventional dose metrics (dose difference (DD), distance‐to‐agreement (DTA), gamma index) show limited sensitivity. The structural similarity index measure (SSIM), combined with radiomics and dosiomics, may better characterize complex dose perturbations. Purpose To evaluate the impact of rotational errors on SBRT dose distributions using SSIM and to develop predictive models integrating radiomic‐dosiomic features. Materials and methods Sixty‐nine clinically validated CyberKnife treatment plans were retrospectively analyzed. In a two‐stage phantom experiment, dose distributions were measured under simulated discrete and non‐discrete rotational errors (ranging from 1.5 to 4.5 degrees), verified by the CyberKnife System Plan QA module. Of the total dose images acquired, 405 datasets (364 from the first stage and 41 from the second stage) passed rigorous quality control. SSIM and gamma passing rates (at 1%/1 mm,1.5%/1.5 mm and 2%/2 mm criteria) were calculated to evaluate their relative sensitivity. Radiomic and dosiomic features were extracted, followed by feature selection using eXtreme Gradient Boosting (XGBoost) with Shapley Additive Explanations (SHAP). Predictive models were optimized via automated machine learning. Performance was assessed with coefficient of determination (R 2 ), mean absolute error (MAE), root mean ‐squared error (RMSE), and median absolute error (MedAE). Results Rotational errors significantly decreased SSIM, showing robust negative correlations with rotation angles across all single‐directional inputs (r ranging from −0.632 to −0.747, all p 0.05), SSIM retained a statistically significant negative correlation ( p < 0.001), demonstrating superior sensitivity. The radiomic‐dosiomic model accurately predicted SSIM (R 2 = 0.818, MAE ≈ 0.0096, RMSE ≈ 0.0123, MedAE ≈ 0.0080). SHAP analysis highlighted wavelet‐transformed dosiomic features as key predictors, and the inclusion of multi‐stage data ensured the model's generalizability across diverse patient characteristics. Conclusions SSIM is a sensitive, structure‐aware metric for evaluating rotational dose perturbations in SBRT, particularly surpassing conventional metrics in detecting complex multi‐directional errors. Integrating radiomic‐dosiomic modeling enhances predictive accuracy and robustness, providing a reliable tool for individualized and automated radiotherapy quality assurance (QA).
Qi et al. (Fri,) studied this question.