R. Wei1, B. Liang2, K. Men3, and J. Dai2; 1National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China, Beijing, China, 2Department of Radiation Oncology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China, 3National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China
Purpose/Objective(s): Intra-fractional tumor motion induced by respiration has a major impact on the accuracy of radiotherapy for liver. Real-time adaptive radiotherapy (RT-AR) may able to eliminate the motion induced negative effect, which aiming at adjusting the treatment plan in real-time according to the actual dose distribution in patient. However, the high complexity of plan optimization makes it hard to balance the effect and efficiency of RT-AR. Thus, we proposed a deep learning based real-time adaptive treatment technique to achieve both effective and efficient RT-AR Materials/
Methods: Three major steps consist of the RT-AR: the organ motion prediction, real-time dose reconstruction and real-time plan optimization. In this study, we utilized three networks to accomplish the above three tasks: a variational encoder-decoder was used to predict 3D organ motion from sequences of 2D cine-MRI images; a U-NET was applied to reconstruct the patient’s actual dose distribution; a multi-branch encoder-decoder was designed to optimize the beam according to the predicted organ motion and reconstructed actual dose distribution in real-time. 35 patients with liver cancer who received MRI-guided radiotherapy were involved in this study. We utilized the single cine-MRI based 3D-MRI reconstruction to achieve the patient anatomy during treatment, and the dose distribution was calculated from the beam optimized with the proposed method. The efficacy of the proposed method was evaluated with the following dosimetry criterion: the percentage of prescribed dose coverage of CTV (Vpres), the mean dose of liver (D_livermean) and the volume of liver whose dose was under 500 cGy(V5). Moreover, we also compared the proposed method with geometry-based MLC tracking method and No motion management situation. Results: For the proposed method, the mean value and standard deviation of Vpres was 96.3%±0.5%, while the geometry based method and No motion management situation had Vpres of 95.8%±0.2% and 92.4%±1.8%. As for the liver, the V5 of the proposed method was 2.3% and 5.1% lower than the geometry based method and No motion management situation. The D_livermean of the proposed method was 1.8% and 7.2% lower than the geometry based method and No motion management situation. The computation time of the proposed method was 134 milliseconds for once optimization. Conclusion: The proposed method achieved effective and efficient real-time adaptive radiotherapy, which showed better target dose delivery and normal tissue protection than conventional geometry-based method as well as no motion management situation
