Research On Deep Learning‐based Dosimetry Evaluation Of Esophageal Radiation Treatment Planning

Intensity modulated radiotherapy(IMRT)is an accurate radiotherapy technique,which is one of major treatments for esophageal cancer.IMRT is a complicated process which aims to maximize radiation at the planned target volume(PTV)while minimize radiation at organs-atrisk(OARs).Dose-volume histogram(DVH)is a practical tool for treatment planner in evaluating plan quality.It provides the fraction of an organ that receives a given amounts of radiation dose and usually presents in a two-dimensional plot.Accurate automatic DVH prediction can provide near-optimal parameters for radiotherapy planning to guide planners to quickly obtain the high-quality treatment plans.Conventional automatic DVH prediction methods usually followed the way of machine learning to define manually feature descriptor for prediction based on the fact that DVH is highly correlated with the geometrical relationship between PTV and OARs.Although these methods have achieved some success,there is still some room for improvement.First,most traditional automated DVH prediction relied on linear dimension reduction techniques to extract geometric and dosimetric features of patients.There is risk in losing certain important non-linear features.Secondly,the calculation process of geometric feature descriptor is more complicated,which makes the model training and testing difficult.Finally,these methods based on manually-defined geometric descriptors demonstrated to be promising in automated DVH prediction,learning to extract the rich spatial relationships between tumors and organs remains challenging.For the above problems,this thesis proposed an automated DVH prediction method based on deep learning to achieve high-quality multi-organ DVH prediction.The main work content of the thesis includes the following points:1)we developed a stacked denoise auto-encoder and a deep belief network to predict OAR’s DVH.Stacked denoising autoencoders can effectively perform non-linear feature reduction on the spatial geometric features between PTV and OARs.The deep belief network models the relationship between patient dosimetric features and spatial geometric features.2)We developed a one-dimensional convolutional neural network to establish the model relationship between the reduced spatial geometric feature vector and the DVH feature vector.The network extracted the local non-linear information of the patient’s geometric features and calculated the relationship with OAR’s DVH.One-dimensional convolutional neural network is more accurate and efficient in training and predicting feature vectors in lower dimensions.3)We developed a spatial geometric-encoding network.The multi-channel distance field images calculated from contour images of PTV and OARs are used as the input of the model.The concurrent spatial and channel squeeze & excitation structure in the proposed model can retain comprehensive spatial information with less computation cost,and achieve the prediction of DVH in complex distribution scenarios of OARs with higher accuracy.4)We have developed a graphical user interface by Py Qt for an automated radiotherapy planning platform for esophageal cancer.The interface embedded a variety of deep learning models,which can perform DICOM-data reader,automatic OAR segmentation,DVH prediction of OARs and three-dimensional dose distribution prediction for esophageal IMRT plans.