Images Segmentation Methods For Glioma Based On Deep Learning

In brain tumors,glioma has a high mortality and disability rate for human beings.Therefore,early detection,diagnosis and treatment can effectively prolong the life span of glioma patients.At the same time,the accurate segmentation of glioma images can play an important role in tumor measurement,evaluation and treatment.With the development of modern medical technology,nuclear magnetic imaging of the brain,and then segmenting the tumor area from it and analyzing it,is an important aid and basis for doctors to diagnose and treat glioma patients.At present,glioma segmentation is still mainly performed manually,that is,the tumor area is segmented by experienced doctors.However,this segmentation method not only requires strong professional knowledge,but also is time-consuming and inefficient.In recent years,deep learning techniques have made breakthroughs in many fields and have been widely used in medical image segmentation.Compared with traditional manual segmentation methods,deep learning methods can learn glioma features autonomously and then realize automatic segmentation,thus,the segmentation methods based on deep learning can effectively improve the efficiency of tumor region segmentation and glioma diagnosis.However,there are limitations and shortcomings in the application of deep learning methods to glioma images segmentation due to the different morphologies and locations of glioma,as well as the complex pathological features,coupled with the serious lack of high-quality glioma annotated images,and the small amount of training sample data will make the generalization of deep learning models insufficient.In response to the above analysis,this paper focuses on two aspects of glioma image samples expansion and precise segmentation of glioma,and the main works are as follows.1.To solve the problem that the lack of glioma image samples affects the segmentation performance of deep learning methods,a glioma images generation method based on multiple discriminator cycle-consistent generative adversarial network(MD-CGAN)is proposed in this paper.First,the proposed MD-CGAN model is used to generate the glioma pathological region images,while the generated glioma pathological region images are then overlaid onto the normal image subregions of the brain to synthesize the glioma images.The objective function of the proposed images generation model includes double adversarial loss and cyclic consistency loss,which avoids the problem of model collapse and ensures the quality and diversity of the brain tumor pathology region images generated by the network.The cyclic consistency loss ensures that MD-CGAN can generate images of glioma pathological regions using images of normal brain regions,while also learning the features of glioma images more comprehensively.2.To address the problem that the diverse morphology and complex pathological features of glioma lead to the low segmentation accuracy of existing deep learning methods,this paper proposes the Feature Pyramid Attention U-Net(FPAU-Net)based on multi-scale feature fusion and attention mechanism for glioma image segmentation.The proposed network is a typical encoding-decoding structure.The left side is the encoding path,which achieves the extraction of glioma images information through a series of convolution and downsampling steps.On the right side is the decoding path,which gradually restores the downsampled feature map to the original size of the input image through a series of upsampling operations.Meanwhile,during the upsampling process of the decoding path,the low-level features and the corresponding high-level features are sent into an attention module,and the fused feature map is obtained after weighting,and then the subsequent upsampling process is carried out.After decoding,the features of each level in the decoding path are fused with multi-scale features,and finally output.The proposed segmentation model FPAU-Net not only introduces the attention mechanism,but also fuses multiple features of different scales,which can effectively improve the segmentation accuracy of brain tumor images.