Research On The Prediction Of 3D Dose Distribution For Intensity Modulated Radiotherapy Based On Deep Learning

Intensity modulated radiation therapy(IMRT)has gradually become the mainstream technology in radiotherapy because it can make the high-dose irradiation area be highly conformal to the target area while protect the organs at risk as much as possible.In the process of designing IMRT plan,physicists often optimize parameters in the treatment planning system based on personal experience and relevant clinical protocols.The entire design process is constantly optimized by manual trial and error.However,this method not only consumes the time and energy of physicists and doctors,but also cannot guarantee whether the final plan is the optimal plan for the current patient.Therefore,in order to assist physicists to quickly formulate high-quality radiotherapy plans and achieve personalized treatment of radiotherapy patients,many scholars have conducted research in the field of dosimetry prediction of radiotherapy plans.Based on the analysis of the pros and cons of the existing dosimetry prediction models,this thesis designs a 3D Res UNet model with attention structure to predict the threedimensional dose distribution of radiotherapy plans,it was named as Dose RAUNet.The model is based on 3D U-net and Res Net,while considering the anatomical structure and field configuration as the input,and it introduces attention mechanism to learn the importance of each field and each anatomical structure.In addition,this thesis designed two comparison models: baseline1 and baseline2,the structure of them are same with the Dose RAUNet but without attention,moreover,the baseline1 only takes anatomical structure as input,and the baseline2 take the anatomical structure and field configuration as input which is different with the baseline1.In this thesis,a total of 110 cases of high-quality cervical cancer IMRT plans were collected for model training and validation.The five-fold cross-validation experiment showed that the predicted and the clinical three-dimensional dose distribution and Dose Volume Histogram(DVH)curves can match well.The average absolute prediction error of all voxels is 0.53％ ±1.44％,the mean absolute dose difference of Planning Target Volume(PTV)and Organ At Risks(OARs)range from 0.21％ to 1.74％.The Dice similarity coefficients(DSC)of different isodose volumes were calculated to analysis the spatial correspondence between the predicted and clinical delivered doses,and the average DSC value was 0.984.Among all the evaluation indicators,the Dos RAUNet shows a higher prediction accuracy and the bsaeline1 has the worst performance.Meanwhile,in order to evaluate the generalization of the Dose RAUNet,32 cases of cervical cancer IMRT plans from another medical institution were collected for model verification.In this batch of test data,the overall mean absolute error of all voxels is 0.96％ ±1.96％,the averaged DSC for all calculated isodose volumes is 0.979.The result shows that the model has a good generalization ability for the same type of plans.At last,in order to verify the prediction accuracy of the model for the plans of non-uniform fields,a total of 88 cases of breast cancer IMRT plans were collected for model training and testing.The results show that the mean absolute error of all voxels is 0.76％ ±3.32％,the averaged DSC for all calculated isodose volumes is 0.961.The mean absolute prediction error for all dosimetric indexes of PTV are less than 2.5％,and the mean absolute prediction error for all dosimetric indexes of OARs are less than 5％,the prediction error can be accepted in clinical.In summary,a three-dimensional dose distribution prediction model for IMRT plans based on Dose RAUNet was proposed in this thesis.This model has a good performance on IMRT plans for different tumors,and it has strong generalization for similar plans.By applying this prediction model to clinical practice,it would be expected to improve the overall plan quality of medical institutions and lay the foundation for automatic planning.