Segmentation Of Organs At Risk In Nasopharyngeal Carinoma Radiotherapy Based On Convolutional Neural Network

Accurate segmentation of organs at risk(OARs)from head and neck(H&N)Computed Tomography(CT)images is crucial for effective nasopharyngeal carcinoma(NPC)radiotherapy.Deep learning techniques involving convolutional neural network(CNN)are currently considered state-of-the-art approaches in detection,classification,and segmentation of images;this approach arises from the recent success in ImageNet Large Scale Visual Recognition Competition and PASCAL VOC detection and segmentation challenges.Though the existing deep learning methods are superior to conventional learning-based methods(that rely on hand-crafted features)in OARs segmentation,most existing models cannot segment all the OARs in a fully end-to-end fashion,as they often design individualized model for each OAR.In addition,when using conventional segmentation network to simultaneously segment all the OARs,the results will be impacted by the similar or neighboring structures and the results often favor big OARs over small OARs.Based on the fact that the general appearance of OARs and surrounding tissues remains similar among CT images of different patients and OARs have a very relatively stable position in the image,and the understanding and analysis of medical image detection and segmentation by deep learning method,this paper is devoted to the study of localization and simultaneous segmentation of various-sized OARs in H&N CT images.This paper mainly includes the following two works:(1)Segmentation of Organs at risk(OARs)in Nasopharyngeal Carcinoma(NPC)based on Detected Target Region.We decomposed the multi-organ segmentation problem of OARs in NPC CT images into OARs localization problem and OARs target region segmentation problem.In the OARs localization problem,we used detection network to predict the location of all OARs in the whole CT image and predict which organ OARs belong to.In the problem of OARs target region segmentation,we took the OARs region image blocks obtained from the detection network as the input of the segmentation network,and obtained the segmentation results of OARs region blocks through iterative depth feature learning.In this approach,segmentation was focused on the detected target area,thereby reducing the influence of similar or neighboring structures on the results.A total of 185 subjects were included in this study for statistical comparison.Sensitivity and specificity were performed to determine the performance of the detection and Dice coefficient was used to quantitatively measure the overlap between automated segmentation results and manual segmentation.Paired-samples t tests and analysis of variance were employed for statistical analysis.Segmentation results from the proposed method correlated strongly with manual segmentation with a mean Dice coefficient(0.861 10.07)for all the OARs.(2)Spatial Aggregation of ROI-Fine-grained Representation Convolutional Neural Network for Organs at risk Joint Localization and Segmentation in Head and Neck CT Images.Our former proposed method predetermined the target organ independently before organ segmentation,causing suboptimal segmentation results as there is no information sharing between the organ localization and segmentation models.Furthermore,most existing models cannot segment all the OARs in a fully end-to-end fashion,as they often design individualized model for each OAR.In addition,when using conventional segmentation network to simultaneously segment all the OARs,the results often favor big OARs over small OARs.Therefore,we propose a deep-learning-based framework performs 3D spatial aggregation and assembling on the proposed multi-organ-segmentation-network-produced multi-OAR probability maps that run on 2D axial,coronal,and sagittal CT planes.In this proposed framework,we propose a multi-organ segmentation network(Seg-FG)to unify OARs localization and segmentation in order to reduce the influence of background region or similar and neighboring structures in input data,and introduce a novel ROI-based fine-grained representation into Seg-FG to efficiently improve the segmentation results of the various-sized OARs.Further,we upgrade the proposed framework with strategy of auto-context model for iteratively improving consistency across slices and also with in-plane pixels.We evaluated our proposed frameworks using two set H&N CT images and achieved preferable and highly robust segmentation performance for OARs of various sizes.