Research On Deep Learning Based Adaptive Radiation Therapy

Radiotherapy is currently one of the most commonly used methods of tumor treatment in clinic,about 70%of cancer patients are treated with radiotherapy either exculsively or in combination with surgery and chemotherapy.The goal of radiotherapy is to ensure the prescription dose delivered to the planning target volume while reduce the radiation dose of organs at risk(OARs),which improves the tumor therapeutic gain.However,the interfraction variations of tumor location and shape,as well as the intrafraction respiratory motion will affect the tumor therapeutic effect.At present,adaptive radiation therapy(ART)is mainly used to solve the above problems,ART modifies the treatment planning according to the interfraction tumor shape and position error feedback which is a continuously optimizing process requiring considerable time and effort.The dose distribution prediction can assist physician to rapidly complete the clinically acceptable plan.However,the dose distribution is not only related to the patient’s anatomical structure,but also affected by the trade-offs between the target dose coverage and OARs sparing.Therefore,the prediction of individualized dose distribution with physician’s preferred trade-offs is an important development trend for ART.During the treatment,ART usually ultilizes CBCT to obtain patient’s anatomical images.However,CBCT cannot be equipped with beam collimator as helical CT,which leads to severe scattering artifacts in CBCT images,restricting the application of CBCT in ART.CBCT is required to rotate around the patient to acquire different angle projections for image reconstruction,while flat-panel X-ray source is a distributed point source imaging device,which can collect multiple angle projections with stationary scanning and reduce the mechanical motion error.It has potential application for the development of new ART imaging system.And constructing the prediction model can evaluate the imaging performance of flat-panel X-ray source in a cost-efficient way,which facilitate optimizing the device structure and dosimetric characteristics.Deep learning has achieved remarkable success in many fields by building neural networks to learn data feature representation.Focusing on the above ART key technical issues,this paper has studied the applications of deep learning method in individualized dose distribution prediction,CBCT scattering correction,and performance prediction model of novel imaging device which is expected to be applied in ART.The main contents of this paper are as follows:1.A deep learning model based on dose volume histogram(DVH)and patient’s anatomy was developed,which can predict individualized dose distribution with physician’s preferred trade-offs.The patient’s anatomy based model can only produce the average conformal dose distribution,cannot achieve the individualized treatment goals for a particular patient.We utilized DVH to quantify the physician’s preferred trade-offs,and the prostate cancer patients were employed to verify the feasibility of the model.2.A scatter distribution estimation method based on Monte Carlo(MC)simulation fused with target distribution modeling via deep reinforcement learning was proposed.The GPU-based MC simulation firstly yielded a raw scatter signal with a low photon number to hasten scatter generation.Then,an optimization objective function integrating Poisson distribution and sparse feature penalty was constructed,and an over-relaxation algorithm was deduced mathematically to solve this objective function.The deep Q-network was built using deep reinforcement learning to interact with the over-relaxation algorithm to intelligently determine the optimal parameters for different cases,so as to obtain the optimal scatter image quality.3.The imaging performance prediction model of the novel device for ART was constructed.This model considered the impacts of the physical characteristics of point source matrix and the geometry configurations of the imaging system.The hyper-parameter in the prediction model was automatically determined by deep reinforcement learning scheme.The experimental results showed that the model can accurately predict the imaging performance of flat-panel X-ray source under different imaging configurations,which contributed to the design and analysis of the flat-panel X-ray source.