摘要
目的探讨基于T2WI、DWI和T1增强序列的2.5D卷积神经网络(convolutional neural network,CNN)在脑膜瘤鉴别诊断中的价值。材料与方法在A、B、C三家医院回顾性收集手术病理证实的脑膜瘤及影像表现与之类似的非脑膜瘤病例共674例,在A医院收集的414例中脑膜瘤为178例,非脑膜瘤为236例,B医院的95例中脑膜瘤为41例,非脑膜瘤为54例,C医院收集的165例中脑膜瘤为78例,非脑膜瘤87例。将所有病例分为5类:孤立性纤维瘤/血管周细胞瘤(Class 0)、脑膜瘤(Class 1)、淋巴瘤(Class 2)、转移瘤(Class 3)、软骨来源及其他类似肿瘤(Class 4)。以A医院队列为训练集,以B医院队列为测试集、C医院队列为验证集,分别基于MRI征象和输入的MRI图像构建梯度决策树(Gradient Boosted Decision Trees,GBDT)模型和三种2.5D CNNs模型(ResNet50、DenseNet169、ResNext50_32x4d)中,在综合比较模型间的性能差异后筛选出最优模型。6位具有不同诊断工作经验的放射医师(初级、中级和高级职称医师各2名)对验证集病例进行独立诊断,评估最优模型与不同经验医师诊断结果的一致性。结果在4种多分类诊断模型中,ResNext50_32x4d模型被判定为最优模型,在训练集、测试集和验证集中的准确度分别为86.7%、82.1%、80.6%;6位具有不同诊断经验的放射医师(医师A~F)在测试集中的准确度分别为61.2%、66.3%、72.1%、77.9%、80.1%、83.2%,最优模型与2位高级职称放射医师的诊断结果具有较好的一致性,组内相关系数(intra-class correlation coefficient,ICC)分别为0.735、0.862。结论基于MRI多序列的2.5D CNN模型在脑膜瘤的鉴别诊断中具有良好的分类预测性能,可为诊断决策提供有价值的参考。
Objective:To explore the value of 2.5D convolutional neural networks(CNN)based on T2WI,DWI,and enhanced T1WI sequences in distinguishing meningiomas from other similar-appearing tumors.Materials and Methods:A total of 674 cases with histopathologically confirmed meningiomas and non-meningiomas with similar imaging features were retrospectively collected from three hospitals(A,B,and C).Among them,414 cases from hospital A(meningiomas,n=178;non-meningiomas,n=236)were used as the training set,95 cases from hospital B(meningiomas,n=41;non-meningiomas,n=54)were used as the test set,and 165 cases from hospital C(meningiomas,n=78;non-meningiomas:n=87)were used as the validation set.All cases were classified into five categories:solitary fibrous tumor/hemangiopericytoma(Class_0),meningioma(Class_1),lymphoma(Class_2),metastatic tumor(Class_3),and cartilage-derived and other similar-appearing tumors(Class_4).A Gradient Boosted Decision Trees(GBDT)model was constructed based on MRI features,and three types of 2.5D CNN,namely ResNet50,DenseNet169,and ResNext50_32x4d,were developed using the input MRI images.After a comprehensive comparison of the performance of these models,the optimal model was selected.Six radiologists with varying levels of experience(two at each level of junior,intermediate,and senior)independently diagnosed cases in the validation set to assess the consistency of the optimal model's diagnostic outcomes with those of radiologists with different levels of experience.Results:Among the four multi-class diagnostic models,ResNext50_32x4d was determined to be the optimal model,with accuracies of 86.7%,82.1%,and 80.6%in the training,test,and validation sets,respectively.Six radiologists with varying levels of diagnostic experience(designed as Radiologist A through Radiologist F)achieved accuracies of 61.2%,66.3%,72.1%,77.9%,80.1%and 83.2%in thevalidation set,respectively.The optimal model showed better consistency with the diagnostic outcomes of the two senior radiologists,with intraclass correlation coefficients(ICC)of 0.735 and 0.862,respectively.Conclusions:The developed 2.5D CNN model based on multi-sequences MRI has good classification and prediction performance in the differential diagnosis of meningiomas,providing valuable reference for distinguishing meningiomas from other brain tumors.
作者
郭开灿
刘婷
刘高元
张勇
刘祥雏
鲁忠燕
周元林
李兵
GUO Kaican;LIU Ting;LIU Gaoyuan;ZHANG Yong;LIU Xiangchu;LU Zhongyan;ZHOU Yuanlin;LI Bing(Department of Radiology,Deyang People's Hospital,Deyang 618000,China;Department of Radiology,Mianzhu People's Hospital,Mianzhu 618200,China;Department of Radiology,Affiliated Hospital of North Sichuan Medical College,Nanchong 637000,China)
出处
《磁共振成像》
北大核心
2025年第2期20-28,共9页
Chinese Journal of Magnetic Resonance Imaging
基金
四川省基层卫生事业发展研究中心重点项目(编号:SWFZ24-Z-01)
德阳市科技计划重点研发项目(编号:2021SZZ079)。
关键词
脑膜瘤
磁共振成像
深度学习
卷积神经网络
鉴别诊断
meningioma
magnetic resonance imaging
deep learning
convolutional neural network
differential diagnosis
