Brain Tumor Segmentation Based On Deep Learning

In recent years,an average of 8.8 million people die of cancer each year,accounting for nearly one-sixth of the total global death toll each year,and the death toll continues to rise.Studies have shown that effective diagnosis can detect cancer early in the disease,making treatment more effective and simpler,and reducing costs.However,the tumors are often hidden in the patient's body tissues.In particular,brain tumors have different shapes and distributions and are difficult to diagnosis,these problems prompted us to explore a solution that can automatically detect and segment tumors and it is highly reliable and helpful for medical diagnosis and treatment.Convolutional neural network has achieved amazing results in the ImageNet visual challenge competition,and has become a focus of current research in the field of computer vision and AI.It has been deeply applied in the field of automatic driving and target recognition,and it has been promoted developed to identify and deals with more difficult tasks.Convolutional neural network is essentially an input-to-output information map.Its purpose is to match the most suitable mappings with a large number of internal parameters,this process can accurately extract the features that we need during task processing.This method effectively solves the problem that it is difficult to improve the accuracy rate caused by insufficient feature extraction in the process of machine learning training,and greatly reduces the influence of manual processing in the model training process.Convolutional neural network also make task processing easier,and the process is truly automated,providing solutions to brain tumor segmentation problems.This paper first systematically analyzes the challenges faced in the study of practical brain tumor segmentation,and summarizes the inherent characteristics of the existing mainstream image semantic segmentation models and the applicability of these models to brain tumor segmentation tasks.Next,an end-to-end segmentation model was proposed for the specificity of the MRI image brain tumor segmentation task.We have innovatively introduced the multi-scale dynamic weighted structure and Refine structure,and explored the impact of these two structures on tumor segmentation tasks.In order to further improve the segmentation accuracy of proposed model,we additionally proposed a Multi-Attention training method.Finally,we discussed the performance of the three structures based on the BRATS 2015 dataset and evaluated the model through the evaluate system provided by the BRATS Experiment Center.The results showed that the proposed model is superior to other methods in complete tumor sections and comparable in the core tumor and enhanced tumor sections,especially,Dice score 0.86 and Sensitivity score 0.91 in the total tumor segmentation obtained the highest score.