Physics‐informed depth‐enhanced surface tracking for cumulative dose reconstruction in breast radiotherapy: An in silico feasibility study
Xixue Tang, Minmin Qiu, Yongjin DengAbstract
Purpose
Surface guided radiotherapy (SGRT) provides nonionizing motion monitoring but lacks direct correlation with internal dosimetric changes. Purely data‐driven models for surface‐to‐internal deformation prediction often yield biomechanically implausible results. This study proposes a physics‐informed depth‐enhanced (PIDE) framework, validated in silico, to reconstruct internal deformation and cumulative dose from noisy surface measurements.
Methods
A virtual clinical trial (VCT) was conducted using weekly cone‐beam CT (CBCT) data from 29 breast cancer patients. We simulated consumer‐grade RGB‐D sensor noise () on extracted patient surfaces. To map external surface features to internal deformation vector fields (DVF), we developed a regularized ridge regression model. Unlike traditional models, ours integrates a spectral energy penalty derived from patient‐specific CBCT motion priors, constraining predictions to physically plausible deformations. Patient‐specific clinical RTDOSE distributions and 348 anatomical landmarks were utilized to evaluate the dosimetric and geometric accuracy. Vector field smoothness was quantified using the Jacobian determinant.
Results
Under realistic simulated sensor noise, the PIDE framework achieved an anatomical target registration error (TRE) of 1.33 ± 1.42 mm (compared to an initial misalignment of 2.14 ± 2.38 mm, representing a 37.8% error reduction). The physics‐constrained regression yielded an RMSE of 2.31 mm ( r = 0.720) for motion prediction and effectively eliminated nonphysical tissue “folding” artifacts (reducing negative Jacobian voxels from 0.54% to 0.10%, p = 0.0209). Dosimetric evaluation using actual patient treatment plans achieved a mean Gamma passing rate (3%/3 mm) of 99.68 ± 0.30% (range: 99.00%–100.00%), with 100% of patients (29/29) exceeding the 90% clinical safety threshold. The end‐to‐end algorithmic inference latency was 2.71 ms per frame, demonstrating high computational efficiency.
Conclusion
The integration of geometric surface features with patient‐specific physical constraints successfully prevents nonphysiological deformation artifacts caused by sensor noise. This in silico study demonstrates the feasibility and computational efficiency of the PIDE framework for rapid cumulative dose monitoring.