Study Of GBM Prognosis Prediction Methods Based On Multi-modal Machine Learning

Gliomas are the most common brain tumor,and glioblastoma multiforme(GBM)is the most aggressive form of gliomas according to the WHO Classification of Tumors of the Central Nervous System.GBM patients account for approximately half of the patients with gliomas,and the overall prognosis of GBM is poor,with most patients living less than two years after initial diagnosis.The median survival time of GBM patients is only 12 to 14 months despite multiple treatments are applied.Furthermore,due to the affection of various factors,GBM patients have diverse prognoses.Therefore,the accurate prognosis prediction of GBM is crucial for choosing a suitable treatment or establishing a hospice care mechanism.Existing studies have shown that multi-modal data such as multi-omics data and whole-slide histopathological images are closely related to GBM prognosis and have great potential in GBM prognosis prediction research.In recent years,with the advance of high-throughput sequencing,whole slide scanning,and other related technologies,many multi-omics and histopathological images are rapidly accumulated.Therefore,how to fuse the above multi-modal data effectively and improve GBM prognosis prediction performance has become the problem that needs to be solved urgently.Therefore.In this dissertation,the status of GBM prognosis prediction methods is deeply investigated,the datasets and pre-process methods commonly used in GBM prognosis prediction are also analyzed.Based on this,the TCGA-GBM dataset,which consists of a variety of multi-modal data including gene expression,DNA methylation,miRNA expression,copy number variation,and histopathological images,is selected as the primary data source.And for the above multi-modal data,through the in-depth study of multi-modal machine learning,novel GBM prognosis prediction methods are proposed for fusing multi-omics data and histopathological images,which improve GBM prognosis performance greatly.The main contributions of this paper are:1.Since most of the existing GBM prognosis prediction methods are based on single-omics data only,the complementary information within multi-omics data is not fully utilized.Thus,a GBM prognosis prediction method based on multiple kernel learning named MO-MKL is proposed.MO-MKL first constructs kernels for each of the omics and then uses multiple kernel learning technology to combine these kernels effectively.Subsequently,MO-MKL fuses gene expression,DNA methylation,copy number variation,and other multi-omics data to predict GBM prognosis.Comparisons between the MO-MKL method and the existing GBM prognosis prediction methods show that MO-MKL can effectively fuse the multi-omics data and increase GBM prognosis prediction performance.2.In addition to multi-omics data,histopathological images are also related to GBM prognosis.Thus,the GBM prognosis prediction method named HI-MKL for fusing multi-omics data and histopathological images is proposed.HI-MKL first extracts histopathological and multi-omics features from histopathological images and multi-omics data respectively,then constructs corresponding kernels.Moreover,to deal with the information loss problem caused by the sparseness of the kernel weights,representative MKL technology is used to combine the above kernels effectively.By analyzing the results of the HI-MKL method,it is shown that HI-MKL can better extract features from multi-modal data.And the information loss problem is resolved while reduces the computational load exceedingly,verifying its superiority in GBM prognosis prediction.3.Based on the above studies,the GBM prognosis prediction method based on multi-modal deep learning by fusing multi-omics data and histopathological images named MDLSurv is proposed.In this dissertation,a novel histopathological image feature extracting sub-neural network is designed,and the multi-omics and histopathological image features are coordinated by deep canonical correlation analysis.Then MDLSurv uses the Cox loss function to train the network.Experiments show that the histopathological image features extracted by the subneural network have stronger predictive power.By fusing the multi-omics and histopathological image features,the prediction performance is significantly improved.