基于组合U-Net网络的CT图像头颈放疗危及器官自动分割

Head and Neck CT Segmentation Based on a Combined U-Net Model

CT图像头颈分割面临着以下难点:CT图像的低对比度导致边界不清,图像扫描间距过大导致冠状面和矢状面图像分辨率低,头颈中待分割的22个器官对于神经网络构建建模的需求不同,且由于存在极小器官造成了类间不平衡。为解决上述问题,该文提出一种U-Net组合模型——由3种U-Net模型组成,分别是2D U-Net模型、3D U-Net模型及3D-small U-Net模型。其中,2D U-Net模型用于厚层图像的分割,3D U-Net模型利用三维空间信息,3D-small U-Net模型用于分割最小的两个器官以解决类不平衡问题。该方法在MICCAI 2019 StructSeg头颈放疗危及器官分割任务中取得了第2名的成绩,平均DSC系数为80.66%,95%豪斯道夫距离为2.96 mm。

Head and neck(HaN)segmentation in CT image is difficult due to low contrast and large slice thickness that resulted in useless information in coronal and sagittal plane for some organs.In addition,complex and small organs have different requirements on neural network modeling.To achieve an accurate segmentation of 22 HaN organs,we combined three U-Net models.The first model was a 2D model that is advantage for dealing with thick slice images.The second model was a 3D model using a cropped input to cover most organs with original resolution in the transverse plane.The third model was a 3D-small U-Net model that focuses on the segmentation of two small organs together and uses a small region of interest(ROI)computed from the bounding box of 2D model segmentation.All the three models were trained using nnUNet method.The final trained model was submitted through docker image to the StructSeg challenge.The leaderboard showed that the proposed method achieved the second place among all methods on ten unseen testing cases with an average Dice value of 80.66%and 95%Hausdorff distance of 2.96 mm.