Research On Brain Glioma Segmentation Method Based On Attention Feature Fusion Convolution Neural Network

Glioma is the most common primary intracranial tumor,which is complicated to treat and has a high mortality rate.Therefore,early diagnosis and treatment are the keys to improving the survival rate of patients.Glioma segmentation is the process of the accurate segmentation of multiple glioma regions from magnetic resonance imaging(MRI),which is very important to glioma diagnosis and treatment.However,manual segmentation of gliomas in MRI is very timeconsuming,and its accuracy depends on professional skills and clinical experience.Therefore,it is vital to propose an efficient and accurate automatic glioma segmentation in clinical diagnosis.In recent years,deep learning algorithms based on convolutional neural networks have shown advanced performance in medical image analysis,providing a new direction for the glioma segmentation.To efficiently and accurately segment gliomas from MRI,this thesis proposes two methods based on attentional feature fusion convolutional neural networks,and two different models are used to automatically segment whole tumor(Whole Tumor,WT),tumor core(Tumor Core,TC)and enhancing tumor(Enhancing Tumor,ET),the specific research contents are as follows:(1)Due to the unclear boundary between glioma and surrounding non-lesional tissue,different sub-regions occupy different positions and proportions of the overall region,making automatic glioma segmentation challenging.This thesis proposes a convolutional neural network based on axis attention and deep supervision to solve the problems during the segment glioma subregions.The Axial Attention Module(AAM)captures long-range dependencies of localglobal feature representations in feature maps by independently performing self-attention mechanisms along three different axes on feature representations,and enhances the modeling ability of long-distance dependence.The ability of the model to identify tumor and non-lesional on fuzzy boundaries improves the accuracy of glioma segmentation.The Deep Supervision(DS)mechanism uses the loss functions at different levels of the decoder to guide the feature extraction during the process of feature map resolution recovery and continuously optimizes the effective feature representation extracted by the early layers,thereby improving the feature expression ability of the model.In addition,this paper also introduces a hybrid loss function to emphasize the lesion area and guide the model to pay more attention to the tumor area,aiming to overcome the problem of identifying tumors of different sizes,thereby suppressing the impact of dataset class imbalance.Finally,the proposed model was tested on the Bra TS 2019 and 2021 datasets,respectively.The DSC values on the Bra TS 2019 dataset were 0.911,0.838,and 0.777 in terms of WT,TC,and ET,respectively.The DSC values on the Bra TS 2021 dataset were0.922,0.861,and 0.830.Experimental results demonstrate that the model proposed in this paper can achieve accurate automatic glioma segmentation to a certain extent.(2)The traditional method fuses different modalities early as network input during the multi-modality MRI glioma image segmentation.However,different modalities share the same weights and perform the same processing,making it impossible to fully extract distinct characteristics.This paper proposes a residual attention convolutional neural network based on a multi-encoder framework to address this issue during the automatic segmentation of gliomas from multi-modality MRI images.The multi-encoder framework can fully extract specific feature representations based on each MRI modality provided by different modalities from multimodality MRI images.Residual Attention(RA)assigns different weights to each MRI modality to focus on the tumor area,thereby reducing the interference between the tumor area and redundant information in normal tissues,and improving the model’s recognition rate of tumors.In addition,the Multi-modality Adaptive Feature Fusion(MAFF)module realizes the fusion of different modal feature representations by adaptively assigning different weights to different modalities,forming richer and more expressive fusion features Indicating that the utilization rate of multimodal features is improved.Finally,the proposed model was tested on the Bra TS 2018 dataset,and the DSC values were 0.908,0.840,and 0.780 in terms of WT,TC,and ET,respectively.The experimental result demonstrates that the model can achieve accurate multimodality MRI glioma segmentation.