BEST IN PHYSICS (MULTI-DISCIPLINARY): Probabilistic Deep Learning-Based Deformable Image Registration with Calibrated Uncertainty Quantification for Adaptive Lung Radiotherapy
Abstract
Purpose
To develop and evaluate a deep-learning-based probabilistic deformable image registration (DIR) framework that provides high-accuracy deformation and calibrated uncertainty, enabling rapid dose accumulation reliability assessment in lung adaptive radiotherapy.
Methods
A variational autoencoder (VAE) was added to the TransMorph DIR model to simultaneously predict deformation fields and voxel-wise aleatoric uncertainty. Two models were trained on intra-institutional datasets: a CT model (725 patients, 1444 CT image phase-pairs) and a low-field MRI-linac model (119 patients, 458 planning-fraction image pairs). Evaluation utilized the public DIRLAB CT dataset (10 patients, 300 landmarks each) and annotated MRI-linac data (16 patients, 107 planning-fraction pairs with 100-400 SIFT-generated landmarks, 7 organ-at-risk (OAR) and target contours each, and for 10 patients, 10 physician-annotated landmarks on one MRI pair per patient). Accuracy was assessed via landmark-based target registration error (TRE) and averaged median Dice similarity coefficient (DSC) between planning and deformed contours across 7 OARs and the GTV. Uncertainty calibration was evaluated using the Expected Calibration Error (ECE) for landmarks and clinical contours. Finally, treatment doses were accumulated to generate dose-volume histograms (DVHs) with uncertainty bands.
Results
Our model achieved mean TRE values of 2.79 mm on DIRLAB and a mean DSC of 0.86 on the MRI-linac dataset. Inference took 1.1 seconds with negligible folding (<0.01%). Registration uncertainty was well-calibrated, yielding landmark ECEs of 1.31 mm (CT), 0.34 mm (MRI SIFT), and 1.56 mm (MRI manual). OAR and GTV contour ECEs were 10% and 8%, respectively. For adaptive cases, predicted PTV D95% deviations within 3% were considered reliable for dose accumulation.
Conclusion
The proposed framework provides a fast, robust method for quantifying registration uncertainty in lung radiotherapy. By integrating calibrated spatial uncertainty into dose accumulation, this approach enables identification of cumulative dose coverage compromised by anatomical variations and registration ambiguity, facilitating safer adaptive treatment decisions.