Prediction Of Radiotherapy Dose Distribution Based On Improved Unet And Similarity Search Of Radiotherapy Plan For Lung Cancer

For different areas of lung cancer patients,automatically make an achieved dose distribution and get radiotherapy plan parameters,aiding physicists to quickly and effectively produce a clinically usable radiotherapy plan.Which can effectively lessen the time for a physicist to make a radiotherapy plan,improve the accuracy of making a radiotherapy plan,and reduce unnecessary injuries to patients’ bodies.Therefore,the use of machine learning algorithms for automated radiotherapy planning has a strong practical application value.This thesis first proposes a framework of dose distribution prediction based on 2D-Unet to predict the dose distribution and meet the requirements of configuration parameter extraction tasks for radiation therapy plan.Further,continuous research had conducted on the problems of low prediction accuracy,low generalization,and inability to directly give configuration parameters of radiotherapy plans.The main contents of this article are as follows:First,a dose distribution prediction framework based 2D-Unet（named as PF-RDDM-2DUnet）is proposed,and simulation experiments are carried out on the radiotherapy planning data of 183 real lung cancer patients.The realization results show that it can initially complete the dose distribution prediction task,but the accuracy is low,and it is not ideal for predicting the dose distribution of the PTV and OAR.At the same time,there are problems such as insufficient generalization under different prescription doses.Second,in order to solve the shortcomings of low accuracy and insufficient generalization of PF-RDDM-2DUnet,a new radiotherapy dose distribution prediction framework（named as PF-RDDM-I3DUnet）is proposed based on improved 3D-Unet.It transforms 2D-Unet into 3D-Unet,and introduces Squeeze-and-Excitation blocks and residual connection to strengthen the model feature extraction ability.At the same time,the mean square error of the predicted dose of PTV and the prescribed dose is added to the original loss function to enhance the prediction ability for dose of PTV.The experimental results show that PF-RDDM-I3DUnet is greatly better than PF-RDDM-2DUnet with the task of dose distribution prediction.Third,in order to solve the problem that PF-RDDM-I3DUnet cannot directly extract the configuration parameters of the radiotherapy plan,a similar search framework（named as SSF_RP₃DRes34RL）of radiotherapy plan is proposed based on 3D-Resnet34 and Triplet-Loss.First,it uses a Triple-Loss network to extract similar feature vectors from ground truth dose distributions,respectively,to construct a feature database.Second,PF-RDDM-I3DUnet is used to predict the dose distribution of a new patient,after that send the predicted dose to the feature extraction network to get it’s feature vector,and then the similarity is calculated between the feature vector of new patient and the feature vector of database.Finally,the radiotherapy plan parameters are extracted from Top-5 similar results.Experimental results show that the recall rate of the SSF_RP₃DRes34RL is about 90%,which can well meet the search tasks of similar radiotherapy plans.