Research On Thoracic Organs Segmentation Algorithms Based On Deep Learning

According to the survey,cancer is one of the causes that threaten people’s health,and the treatment is expensive and the cancer is hard to cure.Radiation therapy is an effective means of treating cancer.The success of radiotherapy depends highly on accurate irradiation to the target and sparing of organs-at-risk(OARs).In order to deliver the prescribed dose to the target volume and reduce the exposure of healthy organs to radiation,the segmentation of high-risk organs is essential for the correct planning of radiation therapy.In clinical practice,segmentation is performed manually by specialists to plan radiation therapy,segmented by doctors is tedious and waste time,and results may vary from specialist to specialist.Automatic segmentation technology provides accurate and robust results,helping physicians analyze the dose delivered to the target volume during radiation therapy in less time.For some reasons,achieving automatic segmentation of high-risk organs is a challenging task such as low-contrast between soft tissue in multi-organ images,variable organ size in different patients,and difference between shapes,so automatic segmentation for multi-organs becomes a important issues in the field of medical images.With the development of deep learning,neural networks perform well in medical image processing and segmentation,gradually replacing manual labeling and semi-automatic segmentation algorithms.The paper is based on the deep learning to research automatic segmentation of multi-organs in the thoracic,committed to solving the problems of low contrast between multiple high-risk organs and surrounding tissues,class imbalance,etc.Using the Seg THOR2019 competition and the Struct Seg2019 public dataset to conduct comparative experiments,the main work includes the following aspects:(1)Aiming at the problems of low contrast between multi-organs and surrounding tissues and noise artifacts in images,a 3D segmentation network for chest multi-organs based on residual transformation and pixel shuffling was proposed.Since the multi-organ medical data has a three-dimensional structure,and the 3DCNN model can make full use of the spatial correlation features,firstly,the fusion residual conversion module replaces the ordinary convolution module to capture the complete spatial background of the organ,and then obtains some images with high resolution for the decoding layer.Then a pixel shuffling module is introduced instead of upsampling in order to further improve the network discrimination ability.Finally,in view of the class imbalance problem of data in multi-organs,the binary cross-entropy classification loss and dice coefficient segmentation loss are combined.(2)Aiming at the problems of low contrast and small size of small organs such as Esophagus and Trachea,it is difficult to identify,and the model training of 3DCNN is timeconsuming and requires high computer performance.However,2DCNN only uses the slice information of the image and ignores the spatial information.A segmentation algorithm fusing context and multi-scale features are proposed.Selecting 2.5D data as network input to obtain slice connections can not only extract 3D contextual information from volume images,but also reduce memory usage and speed up training.In the meanwhile,the efficient global context module is used to capture long-distance dependencies between slice sequences in a single view,and strengthen the connection between channels and spaces.In the encoding layer,the integration of pyramid convolution and dense connection is used to extract multi-scale information.The multi-scale features contain detailed texture information and context information,which enhances the recognition of feature maps by the network.Considering the CT image quality issue,deep supervision is added to refine the output at different scales.(3)Aiming at the problems of each organ has different shape and appearance,and the texture,position and shape vary from patient to patient,and usually requires rich remote context information to accurately segment the organ,a Transformer-based segmentation of multi-organs in the chest is proposed.Although CNNs are able to extract rich features,fully CNN-based methods are insufficient to encode long-range interaction information,either within a single slice or between adjacent slices.Transformer is combined with CNN to capture global dependencies and low-level spatial details,while a fusion module can effectively fuse features with different levels from both branches.