Patient-specific quality assurance, or PSQA, checks whether the dose delivered during radiotherapy agrees with the approved treatment plan. EPID-based methods are attractive because the imaging panel is integrated into the accelerator and can acquire a two-dimensional record of the beam actually emerging from the treatment head. Such images can be compared with calculated portal-dose images and evaluated with gamma analysis, producing both a summary pass rate and a spatial display of discrepancies. The drawback is that conventional EPID PSQA is generally performed once, before treatment begins, so its result describes the machine’s condition on that particular day rather than its condition for every subsequent fraction.

The study addressed that limitation by training a Pix2Pix conditional generative adversarial network to translate an Anisotropic Analytical Algorithm portal-dose calculation into a predicted EPID image. The authors tested two output designs: one that predicted the full EPID intensity and another that predicted the dose-difference map between calculated and measured images. They also compared three input and training configurations: a baseline using only the calculated image, a model conditioned on daily Machine Performance Check data, and a third model that added synthetic multi-leaf collimator deviations to its training set. These axes produced six model variants, all built with the same U-Net-style generator and trained for 2,000 epochs.

The machine-state input consisted of 235 features selected from 253 values recorded by the Halcyon system’s automated daily quality-assurance procedure. Most were related to the positions and backlash of the 57 leaves in each bank of the dual-layer multi-leaf collimator. Seven additional values described beam output, uniformity, field edges, collimator rotation and EPID calibration. Each deviation was normalized by its own quality-control tolerance, allowing features expressed in millimetres, percentages and degrees to enter the network on comparable scales. The vector was projected into a 4096-dimensional embedding, reshaped into a feature map and stacked with the calculated portal-dose image before processing by the U-Net generator.

The dataset comprised 135 normally delivered arc fields from 50 patients treated between April 2023 and October 2024. The patients included eight head-and-neck cases, 34 pelvic cases, of which 26 were prostate treatments, six thorax cases and two extremity cases. All plans used two or three arcs and the 6 MV flattening-filter-free beam. Patient-level separation kept all arcs from a patient in the same subset: 107 fields from 39 patients were used for training and 28 fields from 11 patients formed the common test set. Images were resampled to a uniform 256-by-256 matrix at 1 mm spacing, with no rotation, translation or cropping augmentation.

Because clinically delivered plans rarely contain large collimator errors, the researchers also created 36 deliberately perturbed fields. They selected six base arcs and applied six preset plus-or-minus 1 mm positioning patterns, including asymmetric single-bank shifts, symmetric aperture widening or narrowing, and localized changes affecting selected leaves. The modified plans were physically re-delivered and measured on the accelerator. Corresponding leaf positions in the machine-performance vectors were adjusted by the same offsets, producing paired examples in which the recorded machine state and measured image changed consistently. Twenty-six perturbed fields were added to training, while 10 were combined with their six unperturbed source arcs in a separate plan-paired evaluation set.

On the common 28-field test set, the full difference-target model achieved the best overall agreement with measured gamma passing rates. At 3%/2 mm, its root-mean-square error was 0.48 percentage points and its correlation was 0.944, compared with errors of 0.69 and 0.55 percentage points for the baseline difference-target and EPID-target models. At the stricter 2%/2 mm criterion, the full difference-target model recorded an error of 1.27 percentage points and a correlation of 0.936, while the corresponding baseline errors were 2.01 and 1.68 percentage points. However, bootstrap confidence intervals for the differences between the full and baseline difference-target models included zero, and an exact paired Wilcoxon test found no significant difference in per-field absolute errors on normal plans. The authors therefore characterized the numerical advantage on unperturbed cases as part of the sampling variability rather than confirmed superiority.

The clearest effect appeared in the deliberately perturbed fields. At 3%/2 mm, mean absolute gamma-passing-rate error in the difference-target chain fell from 3.67 percentage points in the baseline model to 2.94 points after adding machine-performance conditioning, then to 2.12 points after also adding deviation-aware augmentation. The full 42% reduction was statistically significant for the paired perturbed samples, with p equal to 0.020. The corresponding reduction in the EPID-target chain was smaller, from 3.45 to 2.71 percentage points, and adding machine-state information alone reversed the direction of that chain’s prediction error. Feature-masking experiments showed why: replacing the 228 MLC-related values with their nominal state increased the full model’s signed error on perturbed plans from positive 1.83 to positive 3.40 percentage points, while masking non-MLC groups had little effect.

Pixel-level detection remained substantially weaker than plan-level prediction. Across all model variants, sensitivity to gamma failures ranged from 19.65% to 32.18%, depending on the criterion and configuration, meaning that most failing pixels were not identified. Specificity and accuracy were high, reaching as much as 99.88% and 99.67%, respectively, but the researchers cautioned that a predicted gamma passing rate cannot replace physical measurement or reliably localize every failure region. The study was also restricted to one institution and one Halcyon v3.0 platform, its perturbation set contained only 1 mm collimator-position offsets, and the plan-paired test set was small. The results therefore establish retrospective agreement and perturbation sensitivity, not clinical readiness, and the authors call for prospective testing of error detection, false-negative risk and stability across treatment fractions.

Subject of Research: Deep learning for machine-state-aware prediction of radiotherapy delivery quality using EPID-based patient-specific quality assurance

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

Article References: Pak, K., Kim, S.-W., Park, K., Cho, B., Kim, Y. J., Song, S. Y., Jeong, C., & Kwak, J. (2026). A deep learning approach for EPID-based patient-specific quality assurance incorporating daily machine-state variations. Frontiers in Oncology, 16, Article 1911466. https://doi.org/10.3389/fonc.2026.1911466

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1911466

Keywords: radiotherapy, patient-specific quality assurance, electronic portal imaging, deep learning, gamma passing rate, linear accelerator, multi-leaf collimator, generative adversarial network, machine performance check, deep, learning, approach