Research On The Prediction Of Respiratory Motion Signals In Dynamic Tumor Radiotherapy

Radiotherapy is an important means of cancer treatment,which can eliminate tumor cells with various types of radiation.Radiation therapy implementation requires close coordination of the localization,planning,verification,and execution stages,striving for complete conformality of the high-dose region and the tumor target area,thus achieving a higher therapeutic gain ratio.However,for thoracic and abdominal tumors,due to the existence of respiratory movement,the radiation exposure position and tumor position will have a large deviation,resulting in an insufficient dose of tumor area and excessive radiation of normal tissue,which will lead to a decline in the treatment effect.The solutions for dynamic tumor radiotherapy include target area enlargement,breath holding,forced shallow breathing,and breath gating technology,but these technologies have certain deficiencies.Adaptive motion compensation real-time tracking technology is the most advanced technology at present,which allows patients to conduct treatment under free breathing without increasing treatment time.It uses an imaging localization system to track the internal marker points or surface substitute points of the tumor in real time to determine the location of the tumor,and adjusts the beam in real time to keep it aimed at the tumor.For real-time tracking radiotherapy systems,it is impossible to achieve instantaneous completion from localization to beam exit,resulting in a certain deviation between the actual irradiation position and the tumor position.Therefore,it is necessary to predict dynamic tumors to compensate for the deviation caused by response delayThis study focuses on the prediction of respiratory motion in dynamic tumor radiotherapy to reduce the deviation caused by response delay.Previous studies modeled and predicted the single-direction respiratory motion signal.In this study,we regard the three-direction respiratory motion signals as a whole,not only considering the temporal pattern of respiratory motion signals but also considering the correlation between different directions.Subsequently,this study proposes a prediction model based on sequential embedding combined with relational embedding and an LGEANet prediction model.Both of them extract temporal pattern feature representation and learn the correlation between different directions of respiratory motion signals by deep learning network model.For model training,this study uses the "pre-training+fine-tuning" mode to obtain the specific prediction model of each patient in a shorter time.At the same time,online fine-tuning is proposed to solve the problem of predicting performance decline caused by changes in respiratory motion patterns during long-term treatment.The deep learning prediction model proposed in this study is essentially a mapping from historical motion signals to the position after the response time,so it can directly predict the internal position after the response delay by external substitute motion signals.This paper uses the respiratory motion signal data provided by a public dataset for experimental analysis.The proposed two models were proved to have relatively better prediction performance through comparison.The 3D direction deviation reduction ratio of each sample predicted for 2-3 minutes is up to 70%on average when the response delay is 150 ms,300 ms,and 450 ms for the sequential embedding combined with the correlation embedding model.The 3D direction deviation reduction ratio of each sample predicted for 15-20 minutes is up to 60%on average when the response delay is 231 ms and 462 ms for the LGEANet model.The online fine-tuning of the LGEANet model has better accuracy than the offline mode in each period.The LGEANet model predicts the internal position after the response delay by substituting the external motion signal,and the maximum 3D direction deviation reduction ratio is 63.11%under the response delay of 450 ms.