Several Application Researches Of Artificial Intelligence In Radiotherapy

Radiation therapy is one of the key techniques for cancer treatment,and above 70%of malignant tumor patients need radiation therapy.The radiation therapy is aim to maximize the gain ratio of radiation therapy,which means focusing the maximized dose on the planning target volume(PTV)to kill the tumor cells and protecting the normal tissue and organs at risk(OARs)from unnecessary irradiation.The radiotherapy treatment procedure includes the following basic steps:(1)Obtain the CT or MR images of the lesions via medical imageing equivement,and delineated the contours of target and OARs;(2)Give the prescription dose,and obtain the treatment planning via Treatment Planning System(TPS);(3)Set-up the patient based on the images scanned before treatment;(4)Do the treatment;If the target region was involved in respiration,tracking the PTV and OARs motion during treatmet is necessary.(5)Evaluate the treatment outcome including the tumor control and raiodamage in OARs.Although the radiotherapy technology is increasing matured,there are many chanllenge problems need bto be solved.Firstly,thecontours of target and OARs are manual delineated by experienced physicians currently.However,the manual contouring is inefficient,and extremely dependent on physician's experience.The large inter-patient variance of manual-delineated contours is also nonnegligible.So,fast automatic segmentation is preferred in clinical.However,the automatic segmentation is not accurate enough to meet the clinical requirements.Moreover,the motion and deformation caused by inspiration would lead to inaccurate dose delivery on target and over radiation on OARs,which will decrease the radiotherapy effects.The accurate respiration motion monitoring is critical for the throax and abdominal cancer radiotherapy.Last but not least,the tumor control and the radioation damage on OARs attract major importance in clinical trials.important section of accurate radiotherapy.Achieving the received dose accumulation for the tumor contraol probability(TCP)and OARs toxicity prediction is meaningful for improving the radiotherapy effects and patient's survival quality.Hence,we constructed researches on OARs automatic segmentation,breath motion monitoring and tracking and radiotherapy outcome prediction.In recent years,artificial intelligence(AI),especially for the machine learning,deep leanring etc.,achieved great successes in areas such as image classification,computer vision.Classical machine learning methods such as principle component analysis,support vector machine,decition tree and random forest are increasingly mature.Deep leanring represented by convolution neural network also show strong ability and is developing fast in image classification and segmentation fields.Many investigations indicate that the great potential of AI in medical fields,especially for disease diagnosis,lesions dectection and segmentation etc..To solve the clinical problems,we have studied on the AI methods including the machine leaning and deep learning to achieve OARs automatic contouring,respotory motion monitoring and tracking,and the O ARs toxicity prediction in radiotherapy.1)Deep learning based OARs automatic segmentation.The convolutional automatic segmentation algorithms cannot perform well in OARs segmentation,especially for OARs with low contrast and small volume,such as the chiasm.In this study,we researched on the segmentation of the right and left eyes,brainstem,right and left optical nerves and chiasm for brain cancer radiotherapy.Considering the relatively stable position relation among OARs,we developed a 3D UNet based recursive segmentation framework.The OARs with good contast in image were contoured firstly.Then smaller region of interest(ROI)covered the other OARs can be extracted based on the position relation among OARs for the further segmentation.For each OAR segmentation,global segmentation was achieved for OAR localization and feature maps(FMs)extraction firstly.On the localized OAR specific ROI,the other model was trained for FMs extraction.A ensemble network combined the global and OAR specific FMs for the final segmentation.MR images from eighty brain cancer patients were used for the segmentation evaluation,including the five-fold cross validation and separated case testing.Mean DC of 94.9%(right eye),94.9%(left eye),90.3%(brainstem),79.6%(right optical nerve),76.9%(left optical nerve)and 68.3%(chiasm)were achieved in the six OARs segmentation testing evaluation,respectively.The results demonstrated the the proposed recursive segmentation framework achievd improvements in the segmentation accuracy and robustness,and decreased the training time cost and complexity.2)Studies of deep learning based respiration monitoring and machine learning based respiration motion tracking.The first one is the deep learning based automatic ROI selection for breast cancer deep inspiration and breath hold(DIBH)radiotherapy.We use deep learning and transfer learning for the registration error(RE)prediction for different ROIs on the thorax-abdominal surface firstly,and then automatically select the optimal ROI based on the predicted RE for accurate respiration state monitoring.The evaluation results on 40 breast patients illustrated that the RMSE/MAE of the selected ROI is smaller than 0.6mm/0.5mm and 0.450/0.35° on translation and rotation RE prediction,respectively.The high RE predictive accuracy demonstrated the feasibility of automaticly ROI selection based on deep learning.The other one is the principle component analysis(PCA)based internal tumor and organs motion tracking for lung cancer radiotherapy.An internal-external motion correlation model was constructed based on the correlation function extracted by PCA between the internal deformation vector fields(DVFs)and externnal DVFs.Then,the PTV and OARs motion and derformation tracking can be achieved by calculating DVFs on the surface.A 4D NCAT phantom and 4DCT images from five lung cancer patients were used for the inter-and inter-fraction tracking accuracy evaluation on the proposed model.The results show that the mean DC of intra-fraction/inter-fraction tumor tracking accuracy on the NCAT phantom and clinical data were 0.96/0.95 and 0.90/0.89,and the mean DC of lung tracking results in all scenarios is 0.97.3)The study of accurate surface dose accumulation algorithm for the application of rectum toxicity prediction for cervical cancer radiotherpay.Considering the close relationship between the accumulated received dose and rectum toxicity,achieving the rectum surface accurate registration to provide accurate DVFs for sdose accumulation is meaningful for the rectum radiation toxicity predction in radiotherapy.Taking the rectum and bladder of cervical patients as targets,we poposed a thin plate spline robuts point matching with local topology preservation algorithm(TPS-RPM-LTP),which is based the local topology should be preseved during the deformation.Based on the TPS-RPM-LTP algorithm,our term achievd the rectum wall dose accumulation and achieved the rectum toxicity prediction based on the machine learning and deep learning models,which demonstrated the feasibility of dose summation and machine learning or deep learning based OARs toxicity prediction.In this study,we studied on several approaches of AI such as the UNet,VGG-16&transfer learning,PCA,SVM,etc.,and applied them in the OARs automatical segmentation for brain cancer raidotherapy,ROI selection for motion monitoring in breast cancer DIBH radiotherapy,PTV and OARs motion tracking in lung cancer radiotherapy and rectum toxicity prediction in cervical cancer radiotherapy,respectively.The good performances achived by AI in these application researches demonstrate the powerfull ability and the extensive application prospect of AI in radiotherapy.