Physics‐informed patient‐specific calibration for monocular visual‐based dose reconstruction in breast cancer radiotherapy
Yongjin Deng, Xuecen Wang, Ruiwan ChenAbstract
Background
Interfractional anatomical variations during breast cancer radiotherapy can significantly deviate the delivered dose from the treatment plan. Standard image guidance via Cone‐Beam CT (CBCT) is typically sparse due to radiation and workflow constraints, leaving daily geometric changes unmonitored.
Purpose
To bridge this gap, we propose a physics‐informed framework that utilizes monocular surface vision for continuous, non‐ionizing 3D dose reconstruction.
Methods
The system utilizes the weekly CBCT scans from the first 2 weeks of treatment (Week 1 and Week 2) for patient‐specific calibration. Surface features are extracted via a hybrid HOG‐CNN descriptor, while 3D deformation vector fields (DVFs) are compressed into a low‐dimensional latent space via Principal Component Analysis (PCA). A linear mapping from surface topography to PCA coefficients is optimized using physics‐informed Tikhonov regularization, incorporating an eigenvalue‐based penalty matrix to suppress high‐order non‐physical modes. Post‐calibration, the model reconstructs 3D dose distributions for subsequent fractions in real time using only monocular input.
Results
Validated on 29 patients (87 fractions), the framework achieved a deformation coefficient correlation of R = 0.796. Reconstructed doses yielded a mean Gamma passing rate (3%/3 mm) of 93.8% ± 3.1% and a dose correlation of 0.940 ± 0.037. Ablation studies demonstrated that removing physics constraints or PCA reduction significantly degraded performance (ΔGPR = −8.5% and −22.1%, respectively; p < 0.001). End‐to‐end latency was 42 ± 5 ms per fraction, representing a ∼124‐fold acceleration over commercial deformable registration solutions.
Conclusions
This framework transforms routine weekly CBCTs into robust “calibration anchors” for continuous visual dose monitoring. It provides clinical‐grade accuracy with negligible computational overhead and no additional ionizing dose, offering a practical solution for real‐time adaptive radiotherapy.