Deep Learning-Based Glioma Segmentation Algorithms For Multi-Modality MR Images

Since glioma is a primary intracranial tumor,accounting for about 81% of brain malignant tumors,and the death rate of malignant glioma is very high,it is important to find and diagnose glioma as soon as possible and formulate an efficient treatment plan to increase the survival rate.Doctors can determine the location and size of glioma by manual labeling on MRI images,but it is time-consuming,labor-consuming and subjective.To improve the consistency,accuracy and efficiency of glioma diagnosis,glioma segmentation algorithm based on deep neural network has attracted more and more attentions.However,the current glioma segmentation methods based on deep neural network still have some problems,such as inhomogeneity intensity of MR image,unclear glioma boundaries,glioma class imbalance and information loss of multi-modality MR images fusion at input DNN layer,etc.In addition,since most of the deep learning models have problems such as too many parameters,high complexity,and low computational efficiency,etc,and medical images occupy a lot of data storage space,it is difficult to deploy the glioma segmentation model based on deep learning into some edge devices to facilitate doctors' usage.Aiming at these image segmentation problems,this paper studies the glioma segmentation algorithm based on deep neural network.The main innovative contributions are listed as follows:(1)Sigmoid-evolution based polarized Cross-Entropy(S-CE)and Hybrid Enhanced-gradient Cross-entropy(HEC)losses are proposed to solve the class-imbalance problem of image segmentation.In order to improve the performance of image segmentation network for glioma region,S-CE loss uses the evolved sigmoid function to obtain loss weights of each pixel,which retains the gradient of incorrectly predicted pixels and suppresses the gradient of correctly predicted pixels.Different from S-CE loss,HEC loss adjusts the cross-entropy loss through exponential function and power function.This adjustment not only suppresses the gradient of correctly predicted pixels,but also strengthens the gradient of incorrectly predicted pixels.(2)A multi-task driven based glioma segmentation method is proposed.This method can solve the problems of blurry boundaries and some discrete wrongly segmented points caused by the invasiveness of glioma and uneven brightness distribution of MRI.According to the hierarchical characteristics of glioma substructure,the glioma network of the proposed method is designed as a progressive segmentation procedure from coarse to fine.Firstly,the whole brain glioma segmentation task achieves the preliminary localization of glioma region,so as to generate the candidate regions for glioma substructure segmentation.Secondly,the features of the whole glioma segmentation task and semantic edge detection task are used as a kind of guidance feature to assist the segmentation of glioma substructure.Finally,the lightweight deep network model is designed by features sharing of different tasks.(3)A glioma segmentation method based on high-and low-frequency fusion and decision level fusion is proposed to solve the problems of noise interference and feature loss in the process of multi-modality MR image fusion.Most of the segmentation methods based on deep networks achieve different multimodal information fusions through simple convolution operations in the input layer.This simple early fusion strategy will lead to problems such as contrast weakening,detail loss and noise interference.In order to solve these problems,firstly,the single-modality high-and low-frequency features are used for the glioma prediction at the first decision level,and then the middle features and the segmentation maps of the first level decision level are fused to obtain the second decision level fusion.Because the second decision level fusion uses a lot of detail features,its segmentation accuracy is higher than the segmentation results from the first level.(4)A low-dimension feature representation network is proposed to alleviate the problem of large storage space occupied by MR images,which leads to a large consumption of storage resources.The storage space occupied by these multi-modality low-dimension features is much smaller than that of the original MRI data.Firstly,as for the loss of detail during multi-modality low-dimension feature fusion,an activity measurement module is proposed to generate the fusion weights of each modality.Then,the normalized weighted fusion module is used to realize content adaptive weighted fusion for the low-dimension feature representation of multi-modality MR images.In order to reduce the feature redundancy of encoding and decoding at multi-resolution,an attention enhanced residual fusion module is proposed.This module selects and removes redundant information of the fused features through spatial and channel attention mechanisms,which efficiently suppresses unimportant features.