通过边缘引导的肾上腺三维CT影像分割  

Edge Guided 3D CT Image Segmentation of Adrenal Gland

作　　者：王文静 牛四杰[1] 李帆 曹彩霞[3] 丛文斌[3] 杨自成 Wang Wenjing;Niu Sijie;Li Fan;Cao Caixia;Cong Wenbin;Yang Zicheng(College of Information Science and Engineering,University of Jinan,Jinan 250000,China;Perception Vision Medical Technologies Co.,Ltd.,Guangzhou 510530,China;The Affiliated Hospital of Qingdao University,Qingdao 266000,China)

机构地区：[1]济南大学信息科学与工程学院,山东济南250000 [2]广州柏视医疗科技有限公司,广东广州510530 [3]青岛大学附属医院,山东青岛266000

出　　处：《南京师大学报(自然科学版)》2025年第1期93-99,共7页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(62101213、62103165、62302191);山东省高等学校人才引育创新团队项目。

摘　　要：计算机断层扫描图像是判断肾脏情况的主要成像方式.医生可以通过分割出腹部CT图像中感兴趣的肾上腺区域,从而计算出肾上腺的体积、灰度值和表面积来判断肾病的病因.然而手工标记图像中的病变区域是耗时、繁琐且具有挑战性的,且病变区域与周围组织极为相似,勾画出的边界极为模糊.因此本文采用一种全卷积神经网络模型MedNeXt——一个受Transformer启发的大核分割网络来对肾上腺3D数据进行体积分割.为应对样本类别不均衡问题,本文还使用对称统一焦点损失替换Dice损失,以提高分割精度.同时考虑到肾上腺组织与周围组织边界难以区分的问题,本文提出结合边界损失函数与主体损失函数同时监督分割过程,使得模型更关注边界的细节信息,从而提升模型性能,实现更精确的分割结果.实验结果表明,所用方法与近几年最新的模型相比在本文所用肾上腺3D数据集上实现了最先进的性能.Computed tomography image is the main imaging method to judge the condition of kidney.Doctors can determine the cause of kidney disease by segmenting the adrenal region of interest in the abdominal CT image and calculating the volume,gray value and surface area of the adrenal gland.However,it is time-consuming,tedious and challenging to manually mark the lesion area in the image,and the lesion area is very similar to the surrounding tissue,and the boundary outlined is extremely fuzzy.Therefore,the method adopted in this paper uses a full convolution neural network model MedNeXt — a transformer inspired large core segmentation network to perform volume segmentation on 3D adrenal data.In order to deal with the problem of unbalanced sample categories,this paper also uses symmetrical unified focus loss to replace Dice loss to improve segmentation accuracy.At the same time,considering the problem that it is difficult to distinguish between adrenal tissue and surrounding tissue boundaries,this paper proposes to combine the boundary loss function and the main body loss function to simultaneously monitor the segmentation process,so that the model pays more attention to the details of the boundary,thus improving the model performance and achieving more accurate segmentation results.Finally,experiments show that the method used in this paper achieves the most advanced performance on the adrenal 3D dataset compared with the latest models in recent years.

关 键 词：全卷积 TRANSFORMER MedNeXt 类别不均衡 体积分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]