Three-dimensional Segmentation And Grading Based On Multimodal Glioma

Glioblastoma is the most common and most deadly primary tumor of the central nervous system.The five-year and 10-year survival rates for patients with low-grade gliomas were 70%and 50%,respectively,with a median survival of about 6.5 to 8 years,whereas the median survival for patients with high-grade gliomas was only 14 months.Because the prognosis of patients with different grades of glioma is very different,doctors need to make corresponding diagnosis and treatment strategies for patients with different grades.In addition,accurate division of edematous area,enhanced tumor area,necrotic area and non-enhanced tumor area and resection of the tumor area as far as possible can greatly improve the survival rate of patients.In view of the small amount of data and the high heterogeneity of imaging phenotype in magnetic resonance images of brain glioma,furthermore,it is difficult to distinguish between different grades of brain gliomas.This paper presents a fast,standard and effective quantitative evaluation method for multi-modality brain gliomas.The first work of this paper combines the local information of the image acquired by the3D convolutional neural network and the advantage of the Transformer encoder in global space modeling to construct a 3D segmentation model of glioma in the whole tumor region of the test set,the mean Dice coefficients of 0.8998,0.8621 and 0.8315 for the core region and the enhanced tumor region were obtained,which were superior to the classical 3D segmentation network.In addition,five kinds of agent tasks are proposed for the model pre-training.The first task is to rotate the image at random and predict the angle;The second task is to segment the image into small blocks and break the order,and finally restore it;The third task is to use x_i~+ as a positive sample and x_i~- as a negative sample,which is different from x_i,the feature vectors z_i?z_i?z_i~+?z_i?z_i~-are fed into the three-variable loss function after the model mapping.The distance between z_i and z_i~+ is reduced,but the distance between z_i and z_i~- is enlarged;The fourth task is to repair a randomly masked image;the fifth task is to use the masked image to predict the Histogram of oriented gradients features of the original image.After supervised learning and fine-tuning the segmentation task,we were able to achieve an average Dice coefficient of 0.8234 in the three tumor regions of the test set with only 30 data-driven cases,in 300 cases of data-driven situation can achieve 0.8866 average Dice,Compared to not using self-supervised learning has a greater improvement.The second work of this paper is based on the segmentation model of the first work,and constructs the model by using the method of imageology to realize the automatic grading of glioma.Based on the four modes of T1,T1Gd,T2 and FLAIR,the whole tumor region and the core region of the tumor were segmented,including first-order statistical features,texture features,wavelet features and morphological features,Pearson correlation coefficient,rank sum test,maximum correlation minimum redundancy and Lasso feature selection algorithm are used to select the optimal feature subset,and then LDA and SVM classifier are used to construct the classification prediction model.The AUC value is 0.9842,the accuracy is 0.955,the specificity is 0.9659,and the specificity is 0.913,the experimental results demonstrate the effectiveness of combining multiple modes and combining multiple tumor subareas.