Abstract
BACKGROUND AND PURPOSE: Electronic portal imaging devices (EPIDs) have been used as an in vivo dosimetry method to identify treatment discrepancies by comparing planned and predicted 3D dose. The purpose of this study was to develop and evaluate a deep learning-based framework for reconstructing three-dimensional (3D) patient dose distributions directly from electronic portal imaging device (EPID) images and planning CT, without relying on Monte Carlo simulation or conventional back-projection methods. MATERIALS AND METHODS: A Res-UNet architecture was trained using 512 beam fields from 60 head and neck IMRT patients to predict beam-specific dose slices from corresponding EPID images and CT data. For each beam, predicted dose slices were assembled into a 3D volume and summed across all beams to reconstruct the full patient dose. The model was evaluated on 10 independent cases using voxel-wise mean absolute error (MAE), mean squared error (MSE), slice-wise dose difference, 3D gamma analysis (global normalization, 10% low-dose threshold) with criteria of 3%/3 mm, and dose-volume histogram (DVH) comparisons. Statistical significance was assessed using the Wilcoxon signed-rank test. RESULTS: The model achieved a mean voxel-wise MAE of 0.163 Gy, with over 95% of voxels with MAE < 2 Gy. The average slice-level dose difference was 1.186 Gy, with localized underestimation observed in high-dose gradient regions. The overall global 3D gamma passing rates were 90.67% for the 3%/3 mm criterion. DVH comparisons showed no statistically significant differences for most targets and organs-at-risk, except for D(2%) of the left lens (p = 0.020) and D(mean) of the parotid glands (p = 0.012), which demonstrated statistically significant differences. The average reconstruction time was 15 s per beam. CONCLUSION: The proposed framework demonstrates the feasibility of fast and patient-specific 3D dose reconstruction using EPID and CT data without complex physical modeling. Under the predefined acceptance criterion (≥95% of voxels with MAE < 2 Gy), 7 out of 10 test cases met the standard. Although high-gradient regions remain challenging and gamma passing rates for the 3%/3 mm criterion were slightly lower than in some prior studies, the method offers sufficient accuracy for many quality assurance applications and provides a foundation for future integration into online adaptive radiotherapy workflows.