A deep learning approach for EPID-based patient-specific quality assurance incorporating daily machine-state variations

Patient-specific quality assurance (PSQA) based on electronic portal imaging devices (EPIDs) is widely used for pretreatment verification, but it is usually performed as a single check of the approved plan and reflects the daily state of the linac only to a limited extent. We propose a deep learning approach that predicts the measured EPID image from the calculated portal dose image using the daily machine state as a conditioning input. A Pix2Pix conditional generative adversarial network was trained to translate Anisotropic Analytical Algorithm portal dose images into measured EPID images, with a 235-dimensional Machine Performance Check (MPC) feature vector added as a conditioning input. Synthetic multi-leaf collimator deviation samples were also introduced during training to represent positioning offsets. Six model variants were evaluated across multiple anatomical sites, combining two prediction targets (the EPID image or the dose-difference map) with three input configurations (baseline, MPC, and MPC with deviation augmentation). Gamma analysis was performed at 3%/3 mm, 3%/2 mm, and 2%/2 mm. The difference-target model with full MPC integration performed best, reaching a root-mean-square error of 0.48 percentage points (pp) and a Pearson correlation of 0.944 for the gamma passing rate (GPR) at 3%/2 mm. On perturbed plans, adding MPC features together with deviation-aware augmentation reduced the absolute GPR prediction error by 42%. Incorporating measured machine-state information improved EPID-based PSQA prediction under the MLC perturbations evaluated in this study. These findings suggest the potential value of a fraction-level, machine-state-aware PSQA, in which the pretreatment measurement establishes a baseline and subsequent daily MPC data are used to update the expected GPR of the approved plan on the day each fraction is delivered. Further prospective evaluation of clinically relevant error detection and false-negative risk is warranted. Nevertheless, this approach may provide a basis for extending PSQA from a single pretreatment assessment toward fraction-level monitoring of patient-specific delivery quality during the treatment course.