Deep Dose Model-based Plan Check And Verification Study For MR-LINACs

The MR-Linacs can provide radiation-free and high soft tissue contrast MR image guidance during patient treatment,and the treatment plan can be adjusted or reoptimized online according to the anatomical changes between or within treatment fractions to achieve online adaptive radiotherapy.Clinical trials have shown that online adaptive radiotherapy has a lot of clinical benefits,and its potential is huge.However,in the MR-Linac workflow,the registration of daily MRI and reference CT,re-contour,electron density assignment,plan adaptation and re-optimization,dose delivery;all these steps are done when the patient is on the couch.It can be seen that,in addition to accuracy,online adaptive radiotherapy has a high demand for efficiency.Due to the introduction of the online plan adjustment mechanism,it is necessary to check the dose of the online plan before dose delivery to prevent major errors in the online plan.In addition,the patient dose will change significantly at tissue heterogeneity boundaries due to the magnetic field dose effect in the MRLinac.For the high-efficiency requirements of online adaptive radiotherapy and the significant influence of magnetic fields on dose distribution,an accurate and efficient secondary dose calculation engine is urgently needed in the MR-Linac online workflow for online plan verification before treatment.Considering that the existing fast dose calculation algorithm cannot consider the magnetic field dose effect,MC simulation can consider the magnetic field effect but its calculation speed doesn’t meet the requirements of adaptive radiotherapy efficiency,in the first work of this study,we developed a millisecond magnetic field dose engine based on the Transformer-CNN model and ray-tracing fluence projection algorithm,which can generate MC-accurate single-beam dose within 310ms,and the average gamma passing rates(2%/2mm,dose threshold 10%)between the predicted dose and MC dose in four sites were greater than 95%.This deep learning-based dose calculation model has the potential to improve the efficiency of online plan verification before treatment,and compared with the existing deep learning-based dose engine in the magnetic field,our method can achieve faster dose calculation speed and a wider range of application sites.The pre-treatment plan check cannot detect errors in dose delivery,during the treatment,whether the actual dose of the patient is equal to the planned dose needs to be verified by comparing the measured dose with the planned dose.Due to the existing scatter correction-based EPID dose reconstruction method cannot consider the magnetic field dose effect,while the dose reconstruction method that calculates the incident fluence first and then uses MC forward calculation can consider the magnetic field effect but does not meet the efficiency requirements,an accurate and efficient dose reconstruction algorithm is urgently needed for online dose verification of MR-Linacs.In the second work of the study,we proposed a two-step dose reconstruction method based on deep learning and EPID back-projection algorithm and achieved fast dose reconstruction using MC-simulated EPID portal dose data and a simple CNN model.The average gamma passing rates(3%/2mm,dose threshold 20%)between the reconstructed dose and MC dose in four sites were greater than 95%,demonstrating the feasibility of proposed method.Compared with the existing scatter correction-based dose reconstruction method in MR-Linacs,this deep learning-based dose reconstruction model can provide an accurate and efficient dose reconstruction scheme when the MR-Linac treats patient sites with lots of heterogeneous regions to improve the accuracy and efficiency of in vivo dose verification.