Physics-Data Fusion Deep Learning Framework for Real-Time X-Ray Acoustic 3D Dose Verification In Radiotherapy
Abstract
Purpose
X-ray-induced acoustic (XA) imaging provides a potential solution for noninvasive, real-time 3D dose verification by characterizing the radiation-induced thermoelastic effect. However, clinical implementation is severely hampered by limited-view sampling and acoustic heterogeneities, which yield ill-posed inverse problems and artifacts in conventional iterative reconstruction. This study proposes a physics-data fusion framework to achieve high-fidelity 3D dose estimation under constrained acquisition conditions.
Methods
We developed a two-stage reconstruction strategy that bridges physical modeling with deep inference. A physically grounded thermoacoustic forward model is first employed to account for heterogeneous wave propagation. The resulting representations are processed by a Cascaded Residual U-Net with Mamba modules (CRU-Mamba). This architecture utilizes convolutional residual blocks for local feature extraction and state-space Mamba modules to model the long-range spatial dependencies inherent in acoustic wavefields, effectively suppressing artifacts arising from limited-view acquisition.
Results
Evaluations on clinical patient datasets demonstrate significant gains in reconstruction fidelity. The proposed framework achieved an average RMSE of 0.022, PSNR of 33.16 dB, and SSIM of 0.984. Critically, Gamma pass rates (3%/3mm) in high-dose regions consistently exceeded 95%. When coupled with adjoint back-projection, the network enables near real-time 3D reconstruction with accuracy comparable to computationally intensive iterative time-reversal algorithms.
Conclusion
By integrating the interpretability of physical modeling with the computational efficiency of Mamba-based architectures, this framework overcomes the trade-off between reconstruction speed and accuracy. It offers a robust and interpretable solution for in vivo 3D dose verification, advancing the feasibility of real-time monitoring in modern radiotherapy.