Modeling dose uncertainty in cone-beam computed tomography: Predictive approach for deep learning-based synthetic computed tomography generation

Cone-beam computed tomography (CBCT) is essential in image-guided radiotherapy (RT) for patient positioning and daily dose calculation. However, CT numbers in CBCT fluctuate and differ from those in computed tomography (CT), requiring synthetic CT (sCT) generation to improve dose calculation accuracy. CBCT-to-sCT synthesis remains a challenging and uncertain task in clinical practice. This study aims to introduce a voxel-wise uncertainty estimator correlated with the error between sCT and CT.

Eighty-five head and neck (H&N) patients treated with photon RT from a single center were selected for developing and validating our uncertainty estimation method. To test the method's robustness on out-of-distribution images, three additional patients from different centers were included. Our proposed uncertainty estimation method builds on established conventional techniques. Additionally, to explore potential error scenarios, we generated several 'plausible' sCTs representing variations in sCT generation caused by CBCT quality differences. This allowed us to quantify dose uncertainties.

The effectiveness of uncertainty maps was evaluated by correlating them with the absolute error map between sCT and CT, yielding a Pearson correlation coefficient between 0.65 and 0.72. Dose uncertainty was determined on the dose-volume histogram (DVH). For all patients except one, the reference CT DVH was included in the uncertainty interval defined by the sCT-derived DVH.

Our proposed methods effectively predict uncertainty maps that aid in evaluating sCT quality. This approach also provides a novel method for estimating dose uncertainty by defining a confidence interval around the CT DVH using the estimated sCT uncertainty.