The Study Of IMRT Automatic Planning Based On The Hybrid Network Of 3D Residuals And U-Net

Intensity Modulated Radiotherapy(IMRT)has been widely used for treating many cancers,being one of the most important radiotherapy technique at present.However,the design process of the inverse IMRT treatment plan is very complex due to a large number of parameters related in particularly to the optimized objective functions.Since the patient-specific dose distributions are unknown beforehand,the objective functions are usually defined by the treatment planners according to standard clinical protocols.Such protocols are based on population-average data without considering individualized dose information.In designing the treatment plans,the planner has to make adjustment repeatedly in a trial-and-error manner until the dose distributions are deemed to meet the clinical specific criteria.However,the experience,skill and time available to the planners vary drastically among different medical centers,resulting in variable treatment plan qualities which seriously affect radiotherapy outcomes of the patients.Therefore,it is extremely necessary to accurately predict the personalized dose distributions and further develop the automatic planning,in order to improve the quality and design efficiency of the treatment plan.At present,the most widely reported approach is the so-called knowledge based radiotherapy(KBRT)which builds a geometry-dosimetry correlation prediction model based on prior patient databases of high-quality treatment plans,with the popular commercial solution of RapidPlan(Varian Medical Systems).However,three obvious limitations exist in the method.Firstly,the method usually relies on the extraction of handcrafted features,while the process is complex and tedious.Meanwhile,handcrafted features from treatment plans do not cover all inherent structure characteristics,so the quality of dose distributions prediction cannot be easily improved.Secondly,most of these predicted dose distributions are expressed as one-dimensional dose volume histogram(DVH)or zero-dimensional dosimetric endpoints that may correspond to non-unique 3D dose distributions.Thirdly,the method usually only predicts the dose distributions,and does not further achieve the automatic planning.The research aim of this thesis is to implement automatic planning of IMRT(including VMAT,and so on).The research tasks included:(1)achieving accurate three dimensional dose distributions prediction for IMRT plans;(2)implementing automatic planning based on accurately predicted dose distributions of IMRT;and finally integrating the automatic planning modules into treatment planning system(TPS)developed by the University of Science and Technology of China(USTC)independently to realize its clinical application.In order to complete the above two research tasks,this research program proposed corresponding materials and methods:firstly,high quality IMRT plans for rectal cancer and VMAT plans for nasopharyngeal cancer were selected from radiotherapy database of the first affiliated hospital of USTC,and the 3D matrices from CT images,structural sets and beam configurations were extracted;The hybrid convolutional neural network named as 3D U-Res-Net was constructed to automatically extract the multi-scale and multi-level features from these 3D matrices to achieve accurate dose distributions prediction.Secondly,a voxel-based optimization model with quadratic loss functions was constructed,and the direct aperture optimization algorithm was employed to realize the automatic planning based on the predicted 3D dose distributions.The prediction accuracy was evaluated in two aspects of 3D dose distributions and dose volume parameters.The qualities of automatic plans were evaluated by the comparison of dose distributions and physical parameters between automatic plans and original clinical plans.The main results were shown as following.(1)The accuracy of dose distributions prediction For rectal cancer IMRT,the average dose prediction bias ranged from(-1.45±5.79)%to(1.58±3.37)%,the mean absolute errors(MAE)varied from(2.63±3.16)%to(5.19±4.78)%;the values of Dice similarity coefficients(DSCs)were above 0.90 for all isodose surfaces,especially being more than 0.95 below 20Gy,and the average DSCs were above 0.92;The differences of all dosimetric parameters were not statistically significant between clinical truth and prediction(P>0.05),except for bladder Dmean.For VMAT plans of nasopharyngeal carcinoma,the average prediction bias ranged from(-2.63±5.72)%to(2.76±5.22)%;the MAEs varied from(2.76±2.90)%to(7.11±7.11)%;The DSCs values were above 0.90 for the isodose surface below 57Gy and above 66Gy;The differences of all dosimetric parameters for OARs and the planned target volumes(PTVs)showed no statistical significance(P>0.05).(2)The results of automatic plans For the rectum IMRT automatic plans,the shape of dose distributions in the about 35Gy dose level and the DVH dosimetric index of most OARs decreased slightly with respect to the original clinical plans,but the difference was not statistically significant(P>0.05);the average number of segments was 76.00±6.94,being higher than original clinical plans,and the average number of Monitor Units was 509.18±108.85,being slightly lower than original clinical plans.For the VMAT automatic plans of nasopharyngeal cancer,the shape of 3D dose distributions and all the dosimetric index of OARs and PTV were similar to the original clinical plan;the average number of Monitor Units for nasopharyngeal cancer VMAT automatic plans was 636.35±108.04,being slightly lower compared to the original clinical plans.For design efficiency of automatic planning,after setting the direction of the beams and click on the optimization of button,the dose prediction module integrated in the DeepPlan system will predict three-dimensional dose distributions in a few seconds,then the dose optimization module will construct optimization functions based on predicted voxel-dose to achieve automatic planning.Therefore,we draw a conclusion that the 3D U-Res-Net model can predict accurately the dose distributions for IMRT and VMAT plan;and the automatic planning can be further achieved based on predicted dose distributions,improving the quality and design efficiency of the treatment plans.