Automatic Parotid Gland Segmentation In MVCT Using Deep Convolutional Neural Networks

Radiation-induced xerostomia,which will seriously affect patients' quality of life,as a major complication in radiation treatment of head and neck cancer,is mainly due to the radiation-induced injury of the parotid gland.Studies show that the parotid gland is one of the important organs at risk in head and neck cancer,and texture features in medical images can be used to predict radiation induced injury in parotid gland.Helical TomoTherapy is equipped with megavoltage CT(MVCT)imaging for image guided radiotherapy and adaptive radiotherapy,as MVCT image contains the body and bony structures and is suitable for alignment verification in radiotherapy.The primary task of parotid gland analysis with MVCT is the segmentation of parotid gland.However,manual segmentation of a large number of MVCT images is timeconsuming,laborious,and subjective.In addition,different from the kilovoltage X-rays diagnostic CT,MVCT includes more Compton effect.Moreover,the parotid gland usually occupies small area in MVCT.In summary,automatic parotid gland segmentation in MVCT is quite challenging due to the parotid gland's high similarity to the surrounding tissue,the amplified doping noise and the low soft tissue contrast in MVCT.Far fewer researches on MVCT segmentation are available in the literature,most of the researches are based on the registration,which relies on the labeled CT images.In this paper,we propose a localization-refinement segmentation scheme based on convolutional neural network(CNN)to segment parotid gland in MVCT.For preprocessing,in order to enhance the quality of MVCT image,we use the BM3D+DFR method to obtain enMVCT.The labels of MVCT are obtained by registration with KVCT images,including rigid registration and deformable registration.After registration,MVCT labels will be checked by experts.In the last step,due to the large background area in original MVCT images,we narrow the gap of the number of pixels between target and background to ensure data-balance of classes.Gray level distribution of pixels in the parotid glands is collected for MVCT image normalization to reduce redundant information and enhance the convergence speed in subsequent model training.In the localization stage,Mask R-CNN is trained with preprocessed MVCT images to acquire regions of interest(ROIs)of the left and right parotid glands.In the refinement stage,an updated variant of U-net model is trained with regions of interest in the training dataset to exactly segment the parotid gland.Modifications of the U-net make the size of the output probability map the same as that of the input image,so that the small input images will not lose information during segmentation.In the implementation of the experiment,35 images from Shandong Cancer Hospital and Institute were acquired for model training and evaluation.We have conducted the contrast experiments of MVCT segmentation on U-net and the proposed scheme.Three data allocation methods were used in each group,MVCT independent testing,enMVCT independent testing and enMVCT mixed testing.In the independent testing,images were assigned to the training set and the testing set according to the difference between patients.In the mixed testing,images were randomly assigned in the same without considering the difference between patients.With respect to MVCT enhancement,the proposed segmentation architecture performs better in enMVCT dataset,and the segmentation performs better in mixed testing.The best segmentation results appear in enMVCT mixed testing of our proposed scheme,with DSC,Hausdorff distance and Jaccard index being 0.841,5.27 and 0.778.Furthermore,to evaluate the proposed scheme in different institutions,we adopt MVCT from Anyang Tumor Hospital,and we can draw a conclusion that the proposed scheme has strong ability of generalization.Finally,based on the localization-refinement scheme,an MVCT automatic segmentation system of parotid gland is established.The main functions of the system include patient information management,MVCT automatic segmentation and feature extraction.The system function,design steps and operation flow are introduced in detail.Moreover,it is proved that the system could automatically segment parotid gland and extract features,providing diagnostic reference for radiologists and meeting the requirements of practical application.