Feasibility of quantifying intra-fractional patient positional error for head and neck IMRT using radiomic features of EPID-measured fluence maps: a phantom study

This study aimed to evaluate the feasibility of estimating patient positional errors during head-and-neck fixed-field intensity-modulated radiation therapy (IMRT) using a machine-learning (ML) regression models based on radiomics features of patient-transmitted fluence maps acquired by an electronic portal imaging device (EPID), as a proof-of-concept toward quantitative intrafractional monitoring. A virtual target, spinal cord, and parotid glands were contoured on an Radiosurgery Verification Phantom simulating nine fixed-field IMRT plans. Treatment plans were created with and without a 0.5 cm or 1.0 cm bolus placed on the phantom surface. For each field, phantom positional errors were simulated in the anterior-posterior (AP), left-right (LR), and superior-inferior (SI) directions at 1 mm intervals from -5 mm to 5 mm. Fluence maps were acquired using an EPID, and difference maps were generated between error and error-free cases. A total of 837 radiomics features were extracted. The dataset included 180 cases (with no bolus or 1.0 cm bolus) for training and 90 cases (with 0.5 cm bolus) for testing. Three ML regression models-random forest, AdaBoost and extra trees-were used. Positional error detection accuracy was evaluated using root mean square error (RMSE), mean absolute error (MAE) and Spearman's rank correlation coefficient. In the test data, RMSE values ranged from 0.550-0.779 mm (AP), 0.440-0.493 mm (LR), and 0.616-0.689 mm (SI). Corresponding MAE values ranged from 0.394-0.584 mm, 0.351-0.392 mm, and 0.492-0.573 mm, respectively. Spearman correlation coefficients ranged from 0.890-0.946 (AP), 0.949-0.953 (LR), and 0.860-0.898 (SI). This phantom-based study suggests that radiomics-based ML regression can quantitatively estimate translational positional-error magnitude from EPID-measured fluence differences under controlled conditions. These findings represent a proof-of-concept for the feasibility of intrafractional geometric monitoring rather than clinical validation. Further investigation using clinically realistic patient data, anatomical changes, and day-to-day machine variations is required before routine clinical implementation.