基于2.5D卷积神经网络模型鉴别高级别胶质瘤与单发脑转移瘤  

Differentiation between high-grade glioma and single brain metastasis based on 2.5D convolutional neural network

作　　者：张斌 黄陈翠 薛彩强 周俊林[2,3,4] Zhang Bin;Huang Chen-cui;Xue Cai-qiang;Zhou Jun-lin(The Second School of Clinical Medicine,Lanzhou 730030,China;Department of Radiology,The Second Hospital of Lanzhou University,Lanzhou 730030,China;Key Laboratory of Medical Imaging of Gansu Province,The Second Hospital of Lanzhou University,Lanzhou 730030,China;Gansu International Scientific and Technological Cooperation Base of Medical Imaging Artificial Intelligence,The Second Hospital of Lanzhou University,Lanzhou 730030,China;Beijing Deepwise&League of PHD Technology Co.,Ltd,Beijing 100080,China)

机构地区：[1]兰州大学第二临床医学院,甘肃兰州730030 [2]兰州大学第二医院放射科,甘肃兰州730030 [3]兰州大学第二医院甘肃省医学影像重点实验室,甘肃兰州730030 [4]兰州大学第二医院医学影像人工智能甘肃省国际科技合作基地,甘肃兰州730030 [5]北京深睿博联科技有限责任公司,北京100080

出　　处：《兰州大学学报（医学版）》2022年第8期23-27,共5页Journal of Lanzhou University（Medical Sciences）

基　　金：国家自然科学基金面上资助项目(82071872);甘肃省科技计划资助项目(21YF5FA123)。

摘　　要：目的探讨基于磁共振成像的2.5D卷积神经网络模型术前鉴别诊断高级别胶质瘤(HGGs)与单发脑转移瘤(BMs)的价值。方法回顾性分析2016年6月-2021年6月兰州大学第二医院经手术病理证实的230例HGGs和111例BMs的T2WI及T1WI对比增强(T1C)图像,预先勾画出2.5D模型下肿瘤区域的感兴趣区,基于ResNet-152卷积神经网络在2.5D维度下分别构建T2WI、T1C及两种序列融合的预测模型,通过受试者操作特征曲线评价模型预测效能。结结果T1C-net模型在训练集和验证集的曲线下面积分别为0.789、0.738,T2-net模型的分别为0.757、0.716,TS-net模型的分别为0.744和0.696。结论基于磁共振成像常规序列的2.5D ResNet模型可以鉴别HGGs和BMs,T1C-net模型性能更好,可以成为鉴别二者并指导临床制定精准化治疗方案的潜在工具。Objective To explore the value of 2.5D convolutional neural network model based on magnetic resonance imaging(MRI)images in the preoperative differential diagnosis of high-grade gliomas(HGGs)and single brain metastasis(BMs).Methods T2WI and T1WI contrast-enhanced(T1C)images of 230 cases of HGGs and 111 cases of BMs confirmed by surgery and pathology in The Second Hospital of Lanzhou University from June 2016 to June 2021 were retrospectively collected,and the region of interest of the tumor area under the 2.5D model was pre-delineated.Based on the ResNet-152 convolutional neural network,T2WI,T1C and two sequence fusion prediction models(T2-net,T1C-net and TS-net)were constructed in the 2.5D dimension,respectively.The predictive performance of the models was evaluated by the receiver operating characteristic curve.Results The AUC values of the T1C-net model in the training and validation sets were0.789 and 0.738,those of the T2-net model were 0.757 and 0.716,and those of the TS-net model were 0.744and 0.696,respectively.Conclusion The 2.5D ResNet model based on MRI conventional sequences could identify HGGs and BMs,and the T1C-net model had a better performance.The model could become a potential tool in identifying two kinds of tumors and guide the clinical formulation of precise treatment plans.

关 键 词：高级别胶质瘤 脑转移瘤 磁共振成像 深度学习 

分 类 号：R445.2[医药卫生—影像医学与核医学] R739.41[医药卫生—诊断学]