摘要
目的对比不同的深度学习分割方法及手动分割方法联合影像组学对高级别胶质瘤(HGG)和单发性脑转移瘤(SBM)患者的MRI图像进行鉴别。方法回顾性分析428例患者的多参数MRI图像,使用Deep Residual U-Net、DenseUnet-based、Fully Convolutional Network、Hybrid Residual U-Net、Nested U-Net及U-Net 6个卷积神经网络模型。首先,在每个序列的图像分割之前对所有图像进行预处理,然后进行训练,训练完成后将模型用来实战分割。然后对这6个模型分割的图像及手动分割的图像进行影像组学特征提取,特征提取完成后进行特征筛选。最后,采用支持向量机、逻辑回归、随机森林及K近邻4种算法建立分类模型。结果采用8种特征组合的方式(6个深度学习卷积网络模型提取出来的特征组合,手动分割提取出来的特征组合及将所有提取出的特征组合到一起的综合组合)评价不同组合下的模型性能,结果显示,在综合组合中的随机森林模型表现最优,其在训练集和测试集上的AUC值分别达到0.95和0.93,展现出优异的分类性能。结论多种深度分割模型联用可以提高鉴别效果,传统手动分割模型优于某些单一深度学习模型。这些模型在区分HGG和SBM疾病方面展现出显著的效能,尤其是结合所有的分割模型提取特征后的综合组合,从而为临床决策和诊疗计划的制定提供了有力的支持。
Objective This study aims to compare different deep learning segmentation methods and manual segmentation techniques combined with radiomics for the differentiation of high-grade glioma(HGG)and solitary brain metastasis(SBM)in patients'magnetic resonance imaging(MRI).Methods A retrospective analysis was conducted on the multiparametric MRI images of 428 patients using six convolutional neural network models:Deep Residual U-Net,DenseUnet-based,Fully Convolutional Network,Hybrid Residual U-Net,Nested U-Net,and U-Net.Initially,all images were preprocessed before image segmentation in each sequence,followed by training.After training,the models were used for practical segmentation.Radiomics features were extracted from the images segmented by these six models and manually segmented images,followed by feature selection.Finally,classification models were built using four algorithms:support vector machine,Logistic regression,random forest,and k-nearest neighbors.Results Eight feature combinations were utilized in this study(features extracted from six deep learning convolutional network models,features extracted from manual segmentation,and a comprehensive combination of all extracted features)to evaluate the performance of models under different combinations.The results showed that the random forest model in the comprehensive combination performed best,achieving AUC values of 0.95 and 0.93 in the training and testing sets,respectively,demonstrating excellent classification performance.ConclusionThe results suggest that the combined use of multiple deep segmentation models can enhance differentiation effectiveness,with traditional manual segmentation models outperforming some individual deep learning models.These models demonstrated significant efficacy in distinguishing between these two diseases,especially when combining features extracted from all segmentation models into a comprehensive combination(unified feature set),thereby providing strong support for clinical decision-making and treatment planning.
作者
徐子超
宗会迁
张娅
柳青
史朝霞
彭兴珍
XU Zichao;ZONG Huiqian;ZHANG Ya(Department of Medical Imaging,The Second Hospital of Hebei Medical University,Shijiazhuang,Hebei Province 050000,P.R.China)
出处
《临床放射学杂志》
2025年第4期591-598,共8页
Journal of Clinical Radiology
基金
河北省医学科学研究课题计划资助项目(编号:20230518)。
关键词
图像分割
高级别胶质瘤
转移瘤
深度学习
影像组学
磁共振成像
Image segmentation
High-grade glioma
Single brain metastases
Deep learning
Radiomics
Magnetic resonance imaging
作者简介
通讯作者:宗会迁。
