The Prognosis Prediction of GBM Based on High-risk Subregion and Multi-parametric MR Imaging

Abstract

Background and object As a heterogeneous tumor, the prognosis information may be different in different region of glioblastoma multiforme (GBM), thus, it's essential to obtain the subregions and evaluate the prognosis of GBM using the high-risk subregions. Based on multi-parametric MR imaging, clearing the purpose of this paper is obtaining the high-risk subregions in GBM and evaluating the GBM prognosis using Multiple Instance Learning (MIL). Methods A total of 104 GBM patients (57 in long-term and 47 in short-term survival group, separated by overall survival of 12 months) were selected from the cancer genome atlas (TCGA), who underwent the T1-weighted contrast-enhanced (T1W-CE), T1-weighted (T1W), T2-weighted (T2W) and FLAIR sequences. In this paper, a region segmentation pattern was presented to characterize the tumor heterogeneity based on machine learning. Combined with four imaging parameters, the GBM region is segmented into multiple subregions by K-means clustering. Extracting the features of co-occurrence matrix, run-length matrix, size zone, and histogram matrix about subregion for building multi-instance bag. A prognostic prediction model was proposed in virtue of the MIL method, using instances to establish a classifier of predicting the high-risk subregions and realizing the prognosis prediction. Finally, the performance of prognosis prediction using instance-based MIL in GBM is presented and the clinical significance is evaluated. Result Training the classifier using discovery cohorts (71 patients including 200 instances) and testing the validation cohorts (33 patients including 94 instances), with the accuracy, sensitivity, and specificity of 81.82%, 76.92%, and 85.00%, respectively. This is superior to our previous results about prognosis prediction with the whole tumor region. Conclusion The prognosis prediction result of our GBM heterogeneity analysis pattern proving the feasibility of the intratumor segmentation using K-means clustering and the effectiveness of Multiple-Instance Learning (MIL) method. It is of benefit to the analysis of GBM high-risk factor, with research value and clinical significance. The intratumor segmentation of the tumor region and the identification of high-risk subregions is benefit to prognosis prediction of GBM, which may be also suitable for other heterogenous tumor.