Study On Pathological Diagnosis Of Glioma And Prediction Of IDH Mutation Based On Semi-supervised Learning

Objective: Glioma classifications are mainly based on the integrated method of histomorphology and molecular genetics currently.However,morphological diagnosis of glioma is subjective.Moreover,there is a shortage of neuropathologists all over the world.Therefore,it is necessary to develop an objective and quantitative tool to improve the accuracy and reliability of glioma diagnosis.Artificial intelligence(AI)is one of powerful tools that can achieve rapidity,repeatability,high accuracy and objective quantity for pathological diagnosis,and can predict biomarkers.This study was on the basis of pathological slide archives of Xiangya hospital using a semi-supervised learning method to train the AI model to identify tumor types among 10 kinds of pathological slides,including normal brain tissue,glial cell hyperplasia tissues,and 8kinds of glioma,and to predict status of IDH mutations.Methods :1.Eight kinds of glioma,normal brain tissue and glial cell proliferation tissue sections were collected;2.All slices were converted into Whole-slide images(WSIs)and upload them to the annotation system;3.WSIs were annotated manually by the initial annotator and reviewed by neuropathologists;4.Glioma classification and IDH genotype prediction AI model were constructed,which was based on convolutional neural network(CNNs)and further verified in the Cancer Genome Atlas(TCGA)dataset.Their performance was compared with neuropathologists.Results:1.The accuracy of the three-category AI model,consisting of astrocytoma,oligodendroglioma and ependymoma,and the five-category AI model,including normal,hyperplasia brain tissue,astrocytoma,oligodendroglioma and ependymoma,was 84.73% and 83.16%respectively;2.The overall accuracy of the AI rating models used for the WHO grade of ependymoma,astrocytoma and oligodendroglioma were 84.07%,80.58% and 77.33%,respectively;3.Our AI model achieved visualization,which was reviewed and approved by neuropathologis;4.The performance of the AI model decreased in the validation experiment of TCGA dataset,but not significantly different from that in our data set;5.The AI model is better than Neuro pathologists in recognition of Oligodendroglioma.However,the grade-rating accuracy of neuropathologists is much higher than that of the AI rating model for WHO grade;6.The prediction accuracy of IDH mutations at Patch level was74.15%.Conclusion: Our AI model could predict IDH mutation,and gain the similar classification performance as neuropathologists,and surpassed in the classification of oligodendroglioma,but still need further improvement in WHO grade of glioma.