A Study On Automatic Detection Of Glioma Lesions And Classification Of Glioma Necrosis And Recurrence

As one of the most commonly primary brain tumors in adults,gliomas grew by infiltrating the surrounding tissue,and could be graded into Low Grade Gliomas and High Grade Gliomas(HGG).HGGs were highly invasive and have a high mortality rate.Clinical treatment of glioma patients are mainly radiotherapy or chemotherapy During radiotherapy,in order to protect the surrounding normal tissue,radiologists need to accurately delineate the tumor area.An accurate and automatic tumor segmentation method can assist doctors in making accurate glioma evaluation and treatment planning.In addition,this radiotherapy therapy usually leads to radiation necrosis,which was the most common side effect in gliomas after treatment.Glioma recurrence and necrosis were difficult to be differentiated by imaging and are treated in completely different ways.Therefore,the clinical classification of glioma recurrence and necrosis is particularly importantRecently,convolutional neural networks(CNNs),including 2D and 3D CNNs,are served as the back-bone in many volumetric image segmentation.However,most of CNN models through consecutive pooling and strided convolutional operations,which leads to the loss of spatial information.In addition,2D convolutions could not fully leverage the spatial information along the third dimension while 3D convolutions suffer from high computational cost and GPU memory consumption.To address these issues,a novel 2D-3D segmentation network was proposed also named Hybrid densely connected network(HD-Net),which consists of a 2D network for efficiently extracting intra-slice features and a 3D counterpart for hierarchically aggregating volumetric contexts under the spirit of the auto-context algorithm for liver and tumor segmentation.We formulated the learning process of HD-Net an end-to-end manner,and captured more high-level intra-slice and inter-slice information by adding Dense Dilated Convolution,Residual Multi-kernel Pooling and Aariational autoencoder modules.Finally,the intra-slice and inter-slice features could be jointly optimized through a fusion feature layerRadiomics has been widely used as a noninvasive method to classify tumor recurrence and necrosis.However,most of the research was based on handcrafted features extracted from the tumor center.The handcrafted features are shallow and low-ordered,could limit the potential of the radiomics model applied.Therefore,we combined deep and handcrafted features extracted from multimodality MRI images to improve the classification accuracy of glioma recurrence versus necrosis.A total of 41,284 handcrafted and 24,576 deep features were extracted for each patient.The 0.623+bootstrap method and the area under the curve(denoted as 0.632+bootstrap AUC)metric were used to select the features.The stepwise forward method was applied to construct 10 logistic regression models based on different combinations of image features.We tested our proposed method on the public available dataset of Multimodal Brain Tumor Segmentation Challenge 2018(BraTS2018)and clinical dataset,respectively.Each sample used T1-weighted postcontrast(TIC),T1-weighted(T1),T2-weighted(T2)and Fluid-attenuated inversion recovery scans(FLIAR)multimodal MRI images.The average Dice value of the proposed HD-Net segmentation network on the 66 samples of BraTS2018 validation set are:whole tumor segmentation(WT,label 0,1,2 and 4)of 0.888,tumor core(TC,label 0,1 and 4)of 0.787 and enhancing tumor(ET,label 0 and 4)of 0.741.We collected 64 gliomas data from the hospital,51 data as the training set and 13 data as the validation set.Considering the limited data set of hospital,we used the pre-trained HD-Net on the 285 samples of BraTS2018 training set to fine-tune the hospital data,and then tested it with the validation set.Mean Dice of WT(label 0,1 and 2)is 0.868 and mean Dice of TC(label 0 and 1)is 0.719.The results on these two data sets demonstrate the effectiveness of our proposed methodFifty-one glioma patients who underwent radiation treatments after surgery were enrolled in this study.Sixteen patients revealed radiation necrosis while 35 patients showed tumor recurrence during the follow-up period.For handcrafted features on multimodality MRI,model 7 with seven features yielded the highest AUC of 0.9624,sensitivity of 0.8497,and specificity of 0.9083 in the validation set.These values were higher than the accuracy of using handcrafted features on single-modality MRI(paired t-test,p<0.05,except sensitivity).For combined handcrafted and AlexNet features on multimodality MRI,model 6 with six features achieved the highest AUC of 0.9982,sensitivity of 0.9941,and specificity of 0.9755 in the validation set.These values were higher than the accuracy of using handcrafted features on multimodality MRI(paired t-test,p<0.05).Handcrafted and deep features extracted from multimodality MRI images reflecting the heterogeneity of gliomas can provide useful information for glioma necrosis/recurrence classification.