Can Enhanced-Couch MPC Replace Off-Isocenter Winston-Lutz QA?
Abstract
Purpose
To test the hypothesis that geometric metrics measured by the enhanced-couch Machine Performance Check (MPC) can serve as a surrogate to predict off-axis isocentricity failure.
Methods
Off-isocenter Winston–Lutz (OIWL) testing was used as the benchmark and was performed with the Sun Nuclear MultiMet-WL phantom and analyzed using MultiMet software (v2.1.0). Enhanced-couch MPC was acquired during the same session. Ten MPC geometric metrics were extracted, including Isocenter_Size, MV/KV_Imager_Projection_Offset, Beam_Center_Shift, Maximum_Positioning_Error, Full_Range_Rotational_Induced_Couch_Shift, Gantry_Absolute/Relative_Offset, Collimator_Rotation_Offset, and Couch_Rotation_Offset. Nine independently calibrated Varian TrueBeam linacs from three hospital systems were included (n=9). A generalized linear model (GLM; binomial logistic regression) modeled binary OIWL outcomes (0=pass, 1=fail). Likelihood-ratio χ² testing guided model selection to identify the most informative predictors. Predictive performance was assessed using leave-one-out cross-validation (LOOCV) ridge-regularized logistic regression, generating out-of-fold probabilities. The classification threshold was tuned to minimizing false negatives and optimizing discrimination, and ROC/AUC was computed.
Results
Two OIWL QAs failure (22.2%) occurred at the largest Full_Range_Rotational_Induced_Couch_Shift values (0.64 and 0.99 mm), while all passed machines had values ≤0.52 mm, demonstrating strong separability in this dataset. The GLM significantly outperformed the intercept-only model (χ²=9.53, p=0.002) and identified Full_Range_Rotational_Induced_Couch_Shift as the dominant predictor. Adding additional predictors did not improve fit (Δχ²=0.000, df=1, p=1.0). This finding was corroborated by feature importance from a bagged-trree classifier. LOOCV ridge-logistic regression achieved Accuracy/Sensitivity/Specificity/Balanced Accuracy = 1.000 with an optimized threshold of 0.230 and out-of-fold AUC=1.000. Out-of-fold probabilities (p_oof) increased monotonically with couch shift; all failures had p_oof >0.228, corresponding to a trigger value of 0.593 mm.
Conclusion
An MPC-derived Full_Range_Rotational_Induced_Couch_Shift metric was highly predictive of OIWL QA failure in this pilot cohort, supporting a trigger-based triage strategy (trigger ~0.6 mm) for selective OIWL testing. Additional validation data will be needed confirm these findings for generalizable clinical criteria.