A Classification Algorithm For IDH Of Glioma Based On Multi-sequence MRI

In the new guidelines for the classification of nervous system tumors issued in 2016,glioma was systematically classified based on IDH genotype.Due to the long detection period and additional trauma in the existing genotype diagnosis technology,the magnetic resonance imaging(MRI)has the advantages of noninvasive,rapid and repeatable,which can be combined with artificial intelligence methods to make medical judgments beyond human power.Therefore,this thesis aims to use multi-sequence MRI technology,based on traditional machine learning and deep learning methods,to study the algorithm of non-invasive diagnosis of glioma IDH genotype.In the study of glioma genotype diagnosis based on traditional machine learning,a glioma dataset was produced,which including CE-T1 W,T2W and ASL sequences,as well as clinical information such as age,gender and IDH phenotype.At the same time,an efficient traditional machine learning modeling process was proposed: first,a total of 851 dimensional radiomics features were extracted from the region of interest of each sequence;then,the radiomics features were combined with age and gender and filtered through Mann Whitney U test,Pearson correlation analysis and LASSO;finally,the SVM models were trained with the remaining features,and the classification performance of single sequence and multi-sequence combination were evaluated.In the study of glioma genotyping diagnosis based on deep learning,firstly,the number of cases and sequences in the self-made dataset were expanded;then,a new three-dimensional deep convolution model was constructed,and the regions of interest of CE-T1 W,T2W,ASL and ADC sequences were directly input into the model,and the age was added in the fully connected stage to improve the model’s performance of;in addition,the classical deep classification model and the traditional machine learning model proposed in this thesis were retrained on the expanded dataset.The classification results based on traditional machine learning show that,compared with any single sequence and feature combination of two sequences,the features from CE-T1 W,T2W and ASL can be used to predict IDH genotypes most accurately;the use of multi-class features can help to distinguish two types of IDH;the use of Pearson correlation analysis can significantly improve the accuracy of prediction;the superiority of this research method was confirmed by comparing with other advanced researches on this dataset.The performance of self-built deep convolution model in IDH genotype diagnosis of glioma is better than that of classical classification mode,and inferior to the traditional machine learning model of this thesis;ADC and clinical features can improve the classification performance of depth model.In short,after training and testing in the same dataset and comparing with other studies,the self-built traditional machine learning and deep learning models can effectively distinguish the two genotypes of IDH according to multi-sequences MRI,but the traditional machine learning model has better classification performance,which can help to improve the level of clinical diagnosis.The problem of over fitting caused by the limited dataset of deep learning method can not be ignored,and there is still room for improvement.Due to the requirement of data preprocessing and the limitation of dataset capacity,the problem of over fitting in deep learning model can not be ignored,and there is still room for improvement.