低场新生儿脑部磁共振图像自动分割方法研究  

Automatic segmentation methods for low-field neonatal brain magnetic resonance images

作　　者：何科毅 陈鸣伸 刘苏锐 周志勇[1,2] 王艳爽 顾艳 戚鑫 李明强 彭博 戴亚康 HE Keyi;CHEN Mingshen;LIU Surui;ZHOU Zhiyong;WANG Yanshuang;GU Yan;QI Xin;LI Mingqiang;PENG Bo;DAI Yakang(Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Sciences,Suzhou 215163,China;School of Biomedical Engineering(Suzhou),Division of Life Sciences and Medicine,University of Science and Technology of China,Suzhou 215163,China;The First Affiliated Hospital of Kangda College of Nanjing Medical University,Lianyungang 222000,China;Jiangsu Lici Medical Equipment Co.,Ltd.,Lianyungang 222000,China)

机构地区：[1]中国科学院苏州生物医学工程技术研究所,苏州215163 [2]中国科学技术大学生命科学与医学部生物医学工程(苏州)学院,苏州215163 [3]南京医科大学康达学院第一附属医院,连云港222000 [4]江苏力磁医疗设备有限公司,连云港22000

出　　处：《中国体视学与图像分析》2024年第4期271-281,共11页Chinese Journal of Stereology and Image Analysis

基　　金：国家自然科学基金项目(62301557,62471467);中国科学院国际杰出学者项目(2025PD0171);中国科学院磁共振技术联盟项目(2024GZL001);江苏省重点研发计划项目(BE2022049-2)。

摘　　要：对于新生儿而言,低场磁共振成像技术是一种更安全、无创的检查技术。然而,这项技术生成的脑部磁共振图像存在对比度不足及图像层厚较大的问题,对其进行分割和分析具有一定的挑战性。在本文中提出了一种快速、准确的低场新生儿脑部磁共振图像自动分割方法,其中包括超分辨率重建、大脑提取和脑组织分割。首先,本文使用N4偏置场校正和最大连通域方法进行数据预处理。其次,实施了一种基于SMORE改进的深度学习方法,用于低场磁共振图像的超分辨率重建,改善了原始SMORE单被试单模型的问题。再次,使用基于错分割损失函数的nnU-Net自配置方法进行大脑提取任务,使得网络更关注错分割区域。最后,采用一种基于体素级对比学习损失的HyperDense-Net方法进行脑组织分割,提升了网络对于组织边缘的分割能力。本文在低场0.35T新生儿脑部磁共振图像上进行训练,在大脑提取任务中Dice系数达到了0.9785,在脑组织分割任务中均值Dice系数达到了0.9405。实验结果表明,本文提出的方法可以有效地提高低场新生儿脑部磁共振图像的分割精度。Low-field magnetic resonance imaging(MRI)technology is a safe and non-invasive examination technique for neonates.However,the brain MRI images generated by this technique have the problems of insufficient contrast and large slice thickness,making segmentation and analysis challenging.In this paper,we propose a fast and accurate method for automatic segmentation of low-field neonate brain MRI images,which includes super-resolution reconstruction,brain extraction,and brain tissue segmentation.Firstly,we preprocess data using N4 bias field correction and the largest connected component method.Then,an improved SMORE-based deep learning method is implemented for super-resolution reconstruction of low-field MRI images,addressing the limitations of the original SMORE single-subject singlemodel.Next,we use the nnU-Net self-configuring method with a mis-segmentation loss function for brain extraction so that the network pays more attention to mis-segmentation regions.Finally,a HyperDense-Net method based on a voxel-level contrastive learning loss is adopted for brain tissue segmentation,which enhances the segmentation ability of the network for tissue edges.In this work,the training was carried out on low-field O.35T magnetic resonance images of neonatal brains.A Dice coefficient of 0.9785 for the brain extraction task and an average Dice coefficient of 0.9405 for the brain tissue segmentation task are achieved.The experimental results demonstrate that the proposed method can effectively improve the segmentation accuracy of low-field magnetic resonance images of neonatal brains.

关 键 词：低场磁共振图像 新生儿 脑组织分割 大脑提取 超分辨率重建 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]