Multimodal CT-MRI Automatic Segmentation Of Small Organs-at-risk Of The Brain Using Cascaded 3D U-Net Network

Accurate and reliable segmentation of Organs-at-risk(OARs)is a prerequisite for radiation treatment planning.Because the traditional manual segmentation is time-consuming and the accuracy is limited by individual experience,methods of automatic segmentation of OARs have gained widespread attention from researchers.However,the current segmentation methods for some small soft tissues in the brain are still not good enough due to small tissue volume and low contrast in CT image datasets.The aim of this M.S.thesis study is to improve automatic segmentation of the small OARs of the brain using multimodal images of CT and MRI in combination with cascaded 3D U-Net algorithms in order to achieve better accuracy than the CT data alone.To achieve the aim of this research,two tasks are performed:(1)Physicians delineate all collected paired CT-MRI data after deformable registration,perform data preprocessing(the removal of the bed board and thermoplastic mask from CT images,the resampling,the deviation of field correction for MRI,and image intensity normalization),generate paired CT-MRI training and test datasets.Build the localization network using the original 3D U-Net,apply it to localization and cropping of the paired CT-MRI images after preprocessing,obtain smaller volume paired CT-MRI images;Based on original 3D U-Net,add the residual difference and deep supervision mechanisms,obtain segmentation network by improving down-sampling,use results from the localization network as the training datasets,yield more accurate segmentation of small OARs of the brain.This study collected 60 cases of the head and neck paired CT-MRI data from the First Affiliated Hospital of Anhui Medical University.All data are preprocessed to generate a training dataset containing of 45 cases of CT-MRI datasets and a test dataset containing 15 cases of CT-MRI data.The cascaded 3D U-Net network is used to train and test the model.The Dic e Similarity Coefficient(DSC)±Standard Deviation(SD)of the left and right eyeball,left and right lens,left and right optic nerve,and optic chiasm are found to be 87.62±2.56,87.54±3.51,74.60 ± 7.90,79.85±3.80,80.39 ±7.12,77.47 ± 9.1 6,respectively.The DSC of the optic nerve,lens and optic chiasm has been improved by(5.4～11.8)%through localization and cropping.The DSC of the optic chiasm is found to be 13.8%higher than the single-modal(CT only)data strategy.The results show that the multi-modal automatic segmentation method proposed in this study can improve the automatic segmentation accuracy of small ORAs of the brain.The small ORAs of the brain,especially the optic chiasm,realize better automatic segmentation owing to the multi-modal data strategy.It is hopeful that the multi-modal automatic segmentation methods will be clinically applied in the future.