基于改进的nnU-Net胰腺分割模型  

Pancreatic segmentation model based on improved nnU-Net

作　　者：龚晓庆[1] GONG Xiaoqing(School of Information Science and Technology,Northwest University,Xi′an 710127,China)

机构地区：[1]西北大学信息科学与技术学院,陕西西安710127

出　　处：《西北大学学报（自然科学版）》2021年第4期594-600,共7页Journal of Northwest University（Natural Science Edition）

基　　金：陕西省重点研发计划资助项目(2021KW-15);国家重点研发计划资助项目子课题(2019YFC1606705-06)。

摘　　要：CT扫描是腹部器官疾病的常规检查手段,准确地对腹部器官进行自动分割能够给医生提供辅助诊疗信息。但腹部CT影像中器官类型多、背景复杂等情况给腹部器官分割带来挑战,尤其是胰腺在CT影像中存在边界模糊的特点,导致现有器官分割模型难以准确分割胰腺。为此,该文针对胰腺分割存在分割边界不准确的问题,基于nnU-Net医学影像分割自适应框架,设计了一种具有边界感知机制的胰腺器官分割模型。该模型在分割网络中嵌入边界感知模块来引导分割网络关注目标边界特征的有效提取;此外,模型将传统分割网络模块提取的语义特征和边界感知模块提取的边界特征进行融合,以有效缓解胰腺器官边界特征提取不完整的问题,从而实现更精准的胰腺器官分割。所提模型在NIH胰腺分割公开数据集上分割准确率达到0.879,分割效果优于现有器官分割模型。CT scan is a routine examination method for abdominal organ diseases,and accurate automatic segmentation of abdominal organs can provide doctors with auxiliary diagnosis and treatment information.However,the multiple types of organs in abdominal CT images and the complex background pose challenges to abdominal organ segmentation.Especially the pancreas has the characteristics of blurred edges in CT images,which makes it difficult for existing organ segmentation models to segment pancreatic organs accurately.For this reason,this paper aims at the problem of inaccurate segmentation edges in pancreatic segmentation.Based on the nnU-Net medical image segmentation adaptive framework,a pancreatic organ segmentation model with edge awareness mechanism is designed.The model embeds an edge awareness module in the segmentation network to guide the segmentation network to focus on the effective extraction of target edge features.In addition,the model fuses the semantic features extracted by the traditional segmentation network module and the edge features extracted by the edge awareness module to effectively relieve the problem of incomplete extraction of the edge features of the pancreas organs,thereby realizing more accurate pancreatic organ segmentation.The proposed model has a segmentation accuracy of 0.879 on the NIH pancreas segmentation public dataset,and the segmentation effect is better than the existing organ segmentation models.

关 键 词：CT影像 胰腺器官分割 nnU-Net分割网络 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]