From Reactive to Predictive: A Digital Twin Framework for Online Proton Prostate SBRT Adaptation
Abstract
Purpose
Online adaptive proton therapy is highly sensitive to interfractional anatomical variation, yet conventional online replanning workflows remain time‑intensive and limit routine clinical implementation, particularly for hypofractionated prostate stereotactic body radiation therapy (SBRT) with a dominant intraprostatic lesion (DIL) boost. This study aims to develop and validate a deep learning-enabled digital twin (DT) framework to enable fast, practical online adaptive proton therapy while preserving plan quality.
Methods
A DT framework integrating deep learning‑based multi‑atlas deformable image registration, corrected daily cone‑beam CT (cCBCT) anatomy update, and knowledge‑based plan quality evaluation was developed. Using an institutional dataset of 43 prostate SBRT patients comprising 210 CBCT volumes, patient‑specific predicted CT (pCT) libraries were generated to represent likely interfractional anatomical variations. For validation, proton treatment plans were pre-planned on the pCT libraries and subsequently selected and rapidly reoptimized online using daily cCBCT. Plan quality was assessed using a ProKnow‑derived composite score evaluating DIL and clinical target volume (CTV) coverage and organ‑at‑risk (OAR) sparing.
Results
DT-guided adaptation achieved clinical‑equivalent or superior plan quality with markedly improved efficiency. The mean online reoptimization time was reduced to 5.50 ± 2.67 minutes, compared with 19.80 ± 11.86 minutes for conventional clinical replanning, corresponding to a 72.22% reduction. Composite plan quality scores were comparable (157.16 ± 5.62 vs. 153.84 ± 6.01). Robust target coverage was maintained with DIL V100 of 99.48% ± 0.59% and CTV V100 of 99.81% ± 0.24%, while OAR sparing for bladder, rectum, and urethra remained within established clinical tolerances.
Conclusion
The proposed DT framework enables fast, clinically feasible online adaptive proton therapy for prostate SBRT with DIL boost. By shifting computational burden to the offline phase and leveraging anatomy‑matched pCT priors, substantial efficiency gains are achieved without compromising dosimetric quality, supporting real‑time personalized proton therapy in routine clinical workflows.