Research On Organs-At-Risk Segmentation Algorithm In Head And Neck CT Images

As a multiple malignant disease,head and neck cancer is highly lethal and difficult to be surgically eradicated.Radiotherapy,as a non-interventional treatment,is a very important means for the treatment of head and neck cancer.According to statistics,about70% of patients will choose radiotherapy to contain the further deterioration of the tumor.The delineation of organs at risk is a very important step in the formulation of radiotherapy plan,which plays a key role in avoiding damage to healthy organs.In head and neck cancer,more than 20 organs at risk need to be delineated in advance.In order to reduce the burden of radiologists,automatic and accurate segmentation of organs at risk has become an urgent task.The main difficulty of this task lies in the complex anatomical structure of the head and neck organs,the fuzzy organ boundary,the imbalance between the size and size of organs and the influence of metal artifacts.In recent years,automatic segmentation of head and neck organs at risk has attracted wide attention and made great progress.However,existing two-stage segmentation methods divide localization and segmentation into two stages,and such algorithms cannot fully share the features of the two stages,which limits the accuracy of segmentation of organs at risk to some extent.Based on this,this paper proposes a dual-branch segmentation network based on fuzzy mask,which fuses localization and segmentation into an end-to-end network,and regresses the fuzzy mask in the localization branch to get the position and shape information of target organs,the segmentation results of target organs are improved by cascading feature maps of the same resolution in two branches to provide certain position and shape guidance for segmentation branch.In addition,in order to make full use of global and local features extracted in the different receptive field,this paper proposes a segmentation network based on information fusion,using the localization branch of large receptive field to extract the global position and shape information of target organs in the whole image,at the same time,the detail information of target organs is extracted with the segmentation branch of the small receptive field,And through the ROI-Crop and the cascade of feature maps,the position and shape information from the localization branch are given guidance to the segmentation branch,makes the global information and local information to fully integrate.This model reduces interference of irrelevant regions on the target organ segmentation while reducing the computing cost,thereby improving the segmentation accuracy of target organs.In order to verify the effectiveness of the proposed methods,the two methods were validated on self-collected dataset and MICCAI Head and Neck Auto Segmentation Challenge 2015 dataset respectively,the results show that the proposed methods can effectively improve the segmentation accuracy of organs at risk.