Application Of Convolutional Neural Network In Grading Of Glioma

Purpose: According to the WHO grading of gliomas in 2016,classification training of preoperative head MRI of patients is running with the application of convolutional neural network.To observe whether the convolutional neural network can be used to analyze the preoperative head MRI,so as to predict the preoperative grading of glioma noninvasively.Method: Collecting preoperative head MRI data from glioma patients who underwent resection of intracranial space-occupying lesions or biopsy in Shengjing Hospital during January 2015 to June 2019.Based on histopathology,they were divided into low grade glioma group(LGG)and high grade glioma group(HGG).All patients met the following criteria:(1)have preoperative head MRI in Shengjing Hospital recently;(2)resection of intracranial space-occupying lesions or biopsy was performed in Shengjing Hospital;(3)The glioma was confirmed pathologically.Data augmentation technique is applied to enlarge the training sets.The adaptive iterative threshold is used to remove the background of head MRI data,after that,all data are standardized.Divide all data into training set,validation set and testing set.They are used to training,adjusting parameters and testing performance of convolutional neural network used in this study.To observe the accuracy of the prediction of glioma grading,and to observe whether the prediction of glioma grading is consistent with the corresponding histopathological grading.Results: The results show that the accuracy,sensitivity and specificity of the convolutional neural network used in this study was 95.0%,91.0% and 93.0%,respectively;the p value of Mc Nemar test was 0.424,and there was no statistical difference between the predicted results and the real results(P > 0.05);the kappa value was 0.860(> 0.70),and the predicted results were in good consistentce with the real results.Conclusions: In this study,convolutional neural network was used to predict the grade of gliomas,which was proved to be highly consistent with postoperative histopathology.It can be used as a tool to assist clinicians in preoperativeevaluation.For the patients considered as gliomas,it is beneficial for the individualized choice of treatment plan to master the grade of glioma accurately,noninvasively and relatively low-cost.For clinicians,automatic and efficient preoperative grade prediction of glioma is beneficial to improve the efficiency of clinical work.