The Research On Medical Image Segmentation Of Brain Tumor Based On Deep Learning

Brain tumor is a persistent disease with high treatment difficulty,high mortality rate and high recurrence rate.At present,among many medical options,surgical resection is the most effective treatment for brain tumor,which can effectively prolong the life of patients.In real life,many patients suffering from brain tumor are not diagnosed timely and accurately,and miss the best time for treatment.Therefore,in early diagnosis,accurate medical image segmentation technology helps doctors to give timely and accurate diagnosis results and carry out subsequent treatment.At present,many medical images assisted manual judgment methods based on deep learning have been proposed,which effectively improve the diagnostic efficiency of doctors.However,due to the characteristics of brain tumors,such as variable morphology,blurred edges of tumor tissues and lack of labeling data,achieving accurate automatic segmentation of brain tumor medical images has become a challenging task in the computer field in recent years.Therefore,this paper aims to achieve accurate segmentation of brain tumor images from the perspective of deep learning algorithms,and the main work includes:(1)To solve the problems of complex tumor morphology and blurred tumor tissue edge segmentation in brain tumor segmentation tasks,this paper proposes a U-Net-based multimodal brain tumor medical image segmentation method for the characteristics of brain tumor data.First,the Rep VGG module is incorporated into the U-Net coding layer to form a multiscale coding module to improve the model’s ability to extract the overall features of the sample.Second,a channel-and space-based attentional feature fusion stitching module SDAM is introduced,which extracts segmented target location information and regional edge information from shallow features by weighted fusion of feature channels and spatial pixels,thereby improving target localization accuracy and alleviating edge blurring problems.Finally,the pyramid module based on subpixel fusion is incorporated into the U-Net decoding layer to alleviate the channel information loss due to up sampling,and thus improve the robustness of the prediction model.It is experimentally demonstrated that each module in the DF-Rep UNet model has validity and the overall performance is better than other classical brain tumor segmentation methods.(2)To solve the problem that many unlabeled brain tumor medical images cannot be effectively utilized,an improved model called MC-MTN is proposed in this paper.The model introduces multi-interpolation consistent regularization and average teacher model,and aims to build a semi-supervised learning model to effectively utilize the features in unlabeled data,to improve the robustness and accuracy of the model when segmenting different samples.First,an improved average teacher network model is proposed to make the model more applicable to medical images of brain tumors by introducing a dynamic decay factor and thus improving the parameter transfer process of the student network and teacher network models.Second,a multi-interpolation consistent regularization method is proposed to enhance the supervised signal in the semi-supervised model by interpolating three images to introduce perturbations in the adversarial direction.It is experimentally demonstrated that the MC-MTN model can effectively enhance the segmentation of DF-Rep UNet on the Bra TS brain tumor dataset,and the overall performance is better than other recent brain tumor segmentation methods.