Direct Dose Correction and Auto-Segmentation on CBCT for Adaptive Radiotherapy with Uncertainty Quantification
Abstract
Purpose
CBCT is essential for daily anatomy assessment in ART, but artifacts and unstable HUs impair dose accuracy, and existing corrections or sCT add uncertainties. This study introduces a unified framework for direct CBCT dose correction and DIR-based auto-segmentation with integrated uncertainty quantification to enable robust ART.
Methods
We proposed an integrated pipeline comprising: (i) three deep learning dose correction models i.e. Dose Adapt CycleGAN (DAC-GAN), Deformable Registration guided Dose GAN (DD-GAN), and Relighting Dose Domain Adaptation Network (RLD-DAN), trained on 98 nasopharyngeal cancer patients to correct CBCT dose to planning CT (pCT) dose; and (ii) two DIR based auto-segmentation networks; Registration guided Segmentation Network (RegSegNet) and Adversarially Regularized Registration guided Segmentation Network (AdvRegSeg) trained on 105 patients to propagate target and OARs onto CBCT. We evaluated performance using voxel-wise image quality metrics, DVH-based dose endpoints, and segmentation metrics. Epistemic uncertainty was estimated using Monte Carlo dropout.
Results
DD-GAN showed the best voxel-level dose agreement (MAE 0.0113) and target coverage (GTV D95: −0.43 Gy; D98: −0.50 Gy), with <1% RMSD versus uncorrected CBCT dose and other models. RLD-DAN achieved the highest structural similarity (SSIM 0.974) and better OAR sparing, with predicted V30 of 33.08 ± 1.92% (right parotid) and 26.68 ± 2.20% (left parotid. For auto-segmentation, RegSegNet provided higher contour accuracy (GTV DSC 0.871 ± 0.053; CTV1 DSC 0.912 ± 0.024) and lower epistemic uncertainty than AdvRegSeg. Monte Carlo dropout supported confidence-aware review by highlighting uncertain regions .
Conclusion
This study demonstrates a practical, end to end framework for dose domain direct CBCT dose correction combined with DIR auto-segmentation, augmented by model uncertainty quantification to support safer adaptive decision making. DD-GAN is best suited for target doses, while RLD-DAN provides more reliable OAR dose behavior together motivating future hybrid/ensemble strategies.