基于改进LeNet-5模型的WHOⅡ/Ⅲ级脑胶质瘤影像自动分级的临床研究

目的探讨基于改进LeNet-5模型的WHOⅡ/Ⅲ级脑胶质瘤影像自动分级的临床应用价值。方法收集经手术病理证实的98例WHOⅡ级和Ⅲ级胶质瘤患者的MRI资料;按照就诊时间顺序将前67例患者作为训练集,后31例患者作为测试集。首先,用深度学习技术及训练集的760张MRI T2WI图像,在卷积神经网络下(LeNet-5模型下)进行训练,模拟影像科医师的判断过程,从而建立了WHOⅡ/Ⅲ级脑胶质瘤影像自动分级系统。然后,对测试集的68张MRI T2WI图像进行临床验证。结果相较于原始LeNet-5模型,改进后的模型对WHOⅡ/Ⅲ级脑胶质瘤的测试准确率明显提高;当epoch=45时,正确率达到最高,epoch大于45时,趋于不变。同时改进后模型的测试错误个数均减少,总体错误率降低;表明改进后的模型对胶质瘤分级的准确率较改进前有提高。结论改进的LeNet-5网络模型能够较好地识别胶质瘤影像表现的深层特征,提高了胶质瘤影像分级的准确率,为临床诊断提供了帮助。

Clinical study on automatic image grading of WHOⅡ/Ⅲglioma based on improved LeNet-5 model

Objective To explore the clinical application value of WHOⅡ/Ⅲglioma image automatic classification based on improved lenet-5 model.Methods The MRI data of 98 patients with WHO gradeⅡandⅢgliomas confirmed by surgery and pathology were collected.The first 67 patients were used as the training set,and the last 31 patients were used as the test set.First of all,760 MRI T2WI images from deep learning technology and training set were trained under convolution neural network(lenet-5 model)to simulate the judgment process of imaging physicians,thus establishing whoⅡ/Ⅲglioma image automatic classification system.Then,68 MRI T2WI images of the test set were clinically verified.Results Compared with the original lenet-5 model,the accuracy of the improved model for who gradeⅡ/Ⅲgliomas was significantly improved.when epoch=45,the accuracy reached the highest.When epoch>45,it tended to remain unchanged.At the same time,the number of test errors and the overall error rate of the improved model were reduced,which indicates that the accuracy of glioma grading of the improved model was improved compared with that of before improvement.Conclusion The improved lenet-5 network model can better identify the deep features of glioma image performance,improve the accuracy of glioma image classification,and provide help for clinical diagnosis.