Deep imaging histology computing predicts the prognosis of GBM patients

Study backgroundglioblastoma (GBM) is a highly invasive growth of malignant brain tumors, belongto in WHO IV gradeCurrently, the most common method of neuroradiology for diagnosing GBM is MRI imaging to determine the location and size of the tumorAccurateidentification and segmentation of multiple abnormal tissues in the tumor in MRI is the basis for predicting patient survivalThe artificial recognition of the tumor size and abnormal tissue within the tumor of MRI imaging can vary from observer to observerZeina AShboul, of the Department of Electrical and Computer Engineering at the University of Virginia, USA, used machine learning algorithms to extract key tumor characteristics from MRI imaging to assess the prognosis of patientsThe results were published in the September 2019 issue of Frontiers in Neuroscienceresearch methods
the authors proposed two-step prediction survival framework using a fully automatic method: the imaging histological characteristics guide deep neural network method for automatic segmentation of tumor tissue, and the classification of total survival regression according to the characteristics of tumor siteMultiple abnormal tissue segmentation steps of tumor can effectively obtain local and all structural characteristics information in MRIThe survival prediction step includes two representative survival prediction models, and selects different characteristics and regression methods to fit (Figure 1)Figure 1Brain tumor imaging group alology structural segmentation flow chartthe study analyzed the BraTS17 training, validation and test data sets, as well as the BraTS18 training, validation, and test data sets, to predict patient survivalThe BraTS17 and BraTS18 data sets contained data on 163 GBM patients, with a total lifetime of days and age in yearsThe training dataset provides data on four patterns, including MRI-T1 weighting, enhanced T1 weighting (T1Gd), T2 weighting (T2) and T2 fluid decay reversal recovery (FLAIR), as well as multiple abnormal tissue segmentations, including enhanced, edema, necrosis, and non-enhanced tissue data The overall lifetime is divided into three groups: long lifetimes greater than 15 months, mid-life of 10 to 15 months, and short-lived periods less than 10 months The authors use the validation data sets of BraTS17 and BraTS18 for verification purposes The BraTS17 validation dataset consists of 33 cases, and the BraTS18 validation dataset consists of 28 cases to predict overall survival The BraTS17 test dataset contains 95 cases, and the BraTS18 test dataset provides 77 cases for testing predictive performance for overall survival all relevant features are extracted from the real lying information of the BraTS18 training data set (Figure 2) The authors found that euler features were selected using recursive feature selection (RFS) and only 39 features were generated out of 28,000 features The 39 Euler features are calculated around the ET profile, 16 are calculated around the WT profile and 7 are calculated around the edema profile The application of RFS in the characteristics produces 23 texture features, 4 histogram features and 8 edema, ET and WT region features respectively The authors use the XGBoost model of leave-one-out cross-validation (leave-one-out cross-validation, LOOCV) for the selected 74 features to predict three corresponding short, medium, and long lifetime categories The results indicate that for the BraTS18 training data set, the classification accuracy is 0.73 (95% CI, 0.655-0.797) Figure 2 Determine the process model for the prognosis The top four key features of the XGBoost model ranking are: the range of tumors on the z-axis, the tumor-strengthening width calculated from the x-axis angle, the contours around edema, and the reinforced tumor The average of the four characteristics significantly stratified 163 patients into low and high risk groups (p value 0.05) (Figure 3) Figure 3 The Kaplan Meier Curve shows the survival of GBM patients 163 GBM patients were divided into two groups: high-risk group (red line) and low risk group (blue line) A Tumor Range; B Tumor enlargement width; C Contour around edema; D Tumor-enhanced contour shaded area represents a 95% confidence interval conclusions , the study proposes a new computational analysis framework, using deep imaging histological calculation methods to make fully automatic segmentation and lifetime prediction of glioblastoma The overall computational analysis framework is designed in two steps, with the first step performing automatic tumor segmentation and the second using the split results for survival prediction Results suggest that accurate segmentation of tumor abnormal tissue types (e.g necrosis, edema, and enhanced tissue) is critical to predicting survival effectiveness.