Research On VMAT Plan Dose Prediction Model Of Cervical Cancer And Its Automatic Planning Based On Machine Learning

Background Cervical cancer is a common malignant tumor disease in female tumor diseases.When designing a treatment plan for cervical cancer patients,the physicist first makes a rough estimate of the Dose volume histogram?DVH?parameters of the organs at risk?OARs?according to his own design experience,and then continuously adjusts the target area and OARs limit parameters.Treatment planning system according to the target parameters for the treatment plan continuous iterative optimization.Due to the large number of OARs,the dose limits of each OAR are strict,and the dose limits of the bladder,rectum,small intestine and femoral head are often too low,resulting in poor dose homogeneity and comformability index in the target area,or dose homogeneity and comformability index are good of the target area,bladder,rectum,small intestine and femoral head dose index exceeds clinical requirements.The physicist needs to constantly perform"trial and error"in the treatment plan until the designed treatment plan meets the clinical requirements,but the treatment plan is usually not optimized,and the quality of the plan is not uniform.The artificial neural network algorithm in machine learning can learn the existing"a priori knowledge"and generate a model,which can be used to complete related tasks.Combination of machine learning and cervical cancer Volumetric modulated arc therapy?VMAT?treatment plan to generate an automatic plan can solve the above problems.The design complexity of the cervical cancer VMAT treatmentplan will be reduced,and the design efficiency and quality will be improved.The application of machine learning in the design of cervical cancer VMAT treatment plan has great potential.Objective The research builds a machine learning based dose prediction model to predict the OARs dosimetry index,and uses the predicted value as an optimization goal to achieve automatic planning.Methods 60 cases of cervical cancer were selected,and the optimization plan was continuously optimized based on the clinical treatment plan.45 optimization plans were selected as the training set,and the remaining 15 optimization plans were used as the verification set.Eight cases in the training set were selected to analyze the dose drop law at the direction of bladder,rectum,small intestine and femoral head,and the spatial anatomical information model of OARs was established.Analysis of the correlation between spatial distance index and bladder,rectum,small intestine dosimetry index(V₃₀,V₄₀,V₅₀),and femoral head dosimetry index(V₃₀,V₃₅,V₄₀),and self-volume.Based on TensorFlow to build artificial neural network,K-fold cross validation method is used to train artificial neural network with OARs spatial distance index,target volume and its corresponding dosimetric index in training set.According to the generalization error,the number of hidden layer nodes is determined.According to the root mean square error,the appropriate artificial neural network was selected as the OARs dosimetry index prediction model,and the dosimetry index prediction is performed on the treatment plan OARs in the verification set.Compare the predicted value with the dosimetry index in the optimization plan to assess the accuracy of the prediction model.Enter the predicted value into the seted automatic plan template to generate an automatic plan.The automatic plan is compared with the clinical treatment plan and the optimization plan to evaluate the advantages and disadvantages of the automatic plan,the clinical treatment plan,and the optimization plan.Results The dose has a different drop rate around each OARs.The correlation between spatial distance index and bladder,rectum,small intestine dosimetry index(V₃₀,V₄₀,V₅₀),and femoral head dosimetry index(V₃₀,V₃₅,V₄₀),and self-volume of the majority were moderately correlated and strongly correlated.The root mean square errors of the dosimetric index models for bladder,rectum,small intestine and femoral head were18.74?%?,15.86?%?,1.53?%?,and 1.71?%?,respectively,and the hidden layer nodes were12,20,10,and 10,respectively.OARs dosimetry index prediction model can better predict bladder,rectum,small intestine dosimetry index(V₃₀,V₄₀,V₅₀)and femoral head dosimetry index(V₃₀,V₃₅,V₄₀),The difference between predictive value and optimization plan dosimetry index was no statistically significant.Dose homogeneity index was no significant difference between the automatic plan and the clinical treatment plan,and optimization plan,but the dose conformity was increased by 5.0%and 4.9%,respectively?Z=-2.05,2.16,P<0.05?,compared with the clinical treatment plan and optimization plan.The mean value of most OARs dosimetry index in the automatic plan was lower than the clinical treatment plan,and the bladder V₅₀ index in the automatic plan was 2.95?%?lower than the clinical plan?t=-2.912,P<0.05?.Most of the OARs dosimetry index is similar to the optimization plan in the automatic plan.But the V₅₀ dosimetry index of the small intestine and the V₃₅ and V₄₀ dose index of the femoral head are 0.84?%?,2.67?%?,and 1.96?%?lower than the optimization plan,respectively?Z=-1.761,t=-3.93,Z=-3.51,P<0.05?.Conclusion The dose index prediction model based on artificial neural network of machine learning can accurately predict the dosimetric index of OARs.The OARs dosimetry index,target dose homogeneity and comformability index in the automatic planning are superior to the clinical treatment plan and optimization plan,and meet the clinical treatment plan design requirements.It is feasible to combine the machine learning-based OARs dose index prediction model with the planning template to generate a cervical cancer VMAT automatic plan.