An uncertainty-aware deep learning-based registration model for adaptive lung cancer radiotherapy

OBJECTIVE: To develop a deep learning (DL) based probabilistic 3D deformable image registration framework for uncertainty-aware dose accumulation in magnetic resonance imaging-guided radiotherapy (MRgRT) of lung cancer. Approach: We propose TransMorph-prob (TM-prob), a variational Bayesian model that predicts the voxel-wise mean and variance of dense displacement fields (DDFs). Two TM-prob models were trained on an internal CT dataset (725 patients) and a low-field MR-linac dataset (102 patients), and benchmarked against Monte Carlo dropout TransMorph (TM-MCD) and DIRUncertainty with B-spline initialization (ASDIR) on the DIRLAB 4DCT dataset (10 patients, 3000 landmarks), and an annotated MR-linac dataset (16 patients, 107 image pairs, 227 landmarks) with target and seven organ-at-risk (OAR) contours. Both datasets contained an additional 1239/13,289 algorithm-generated landmarks. Performance was evaluated using target registration error (TRE), Dice similarity coefficient (DSC), landmark- and voxel-wise expected calibration errors (LECE/ECE), and dose distributions derived from 2000 DDF samples per patient. Main results: On the DIRLAB dataset, TM-prob achieved a median TRE of \qty{1.8}{\mm}/\qty{2.8}{\mm} on manual/algorithm-generated landmarks, versus \qty{1.3}{\mm}/\qty{2.4}{\mm} for ASDIR and \qty{2.1}{\mm}/\qty{3.0}{\mm} for TM-MCD. On the MRI dataset, TM-prob achieved the median TRE of \qty{3.9}{\mm}/\qty{2.6}{\mm} and a mean DSC of 0.87, matching ASDIR (\qty{3.3}{\mm}/\qty{2.7}{\mm}, and DSC of 0.87) while outperforming TM-MCD (\qty{4.4}{\mm}/\qty{2.9}{\mm}, and 0.84). In uncertainty calibration, TM-prob achieved a LECE of \qty{1.74}{\mm}/\qty{0.30}{\mm} on MRI manual/algorithm-generated landmarks versus \qty{2.41}{\mm}/\qty{0.56}{\mm} for ASDIR and \qty{6.65}{\mm}/\qty{3.92}{\mm} for TM-MCD, and an average contour ECE of 5.8% versus 7.0% and 21.2%. Two patients with indistinct GTV boundaries exhibited substantially wider target D95 distributions, demonstrating that calibrated DDF uncertainty translates into clinically relevant dose-coverage uncertainty. Significance: TM-prob is an end-to-end DL model, achieving fast, well-calibrated, uncertainty-aware registration for lung cancer patients, enabling reliable DDF uncertainty propagation into dose accumulation for clinical decision-making in online adaptive radiotherapy.