Fast fourier transform based spectral features for machine learning prediction of gamma passing rates in virtual VMAT...

This study aimed to evaluate the feasibility of fast Fourier transform (FFT) based spectral features for predicting volumetric modulated arc therapy (VMAT) gamma passing rates (GPRs) in virtual patient-specific quality assurance (PSQA).

A total of 481 VMAT treatment plans were retrospectively collected. Multileaf collimator (MLC) trajectories and monitor units (MU) were extracted from control points (CPs) and interpolated onto a uniform gantry-angle grid. FFT-based power spectra were then computed to extract global, spatial-axis, and angular-axis spectral features characterizing modulation complexity. These spectral features, alongside 25 conventional plan complexity metrics, were used to train Random Forest (RF) and XGBoost models for GPR regression and PSQA pass/fail classification. We employed a nested cross-validation strategy and used SHapley Additive exPlanations (SHAP) for model interpretation.

Spectral features and conventional plan complexity metrics correlated significantly with GPRs, particularly spatial-axis high-frequency energy ratios and total spectral energy. In the regression task, the XGBoost model utilizing the combined feature set achieved the lowest mean absolute error (MAE) of 1.095 ± 0.182. For classification, the RF model with combined features yielded an AUC of 0.886 ± 0.043, an accuracy of 0.844 ± 0.066, and an F1 score of 0.602 ± 0.094. SHAP analysis confirmed that spectral energy, spatial-axis low- and high-frequency energy ratios, were major contributors to predicting PSQA failures.

FFT-based spectral analysis provides physically interpretable features for characterizing VMAT modulation complexity. These features showed feasible performance for GPR prediction and PSQA classification. These findings support their potential utility as interpretable inputs for virtual VMAT PSQA.