Research Of Brain Tumor Segmentation Based On Deep Learning

Artificial intelligence and computer vision have developed rapidly in recent years and become an important part of the Internet technology industry.Deep learning algorithms,especially convolutional neural networks,stand out among artificial intelligence algorithms.For the analysis of complex data sets,the effect of deep learning has been far ahead of other traditional algorithms.As the main research content of computer vision,image segmentation has been applied in many fields,such as intelligent medical,autonomous driving,indoor navigation,human-computer interaction,virtual or augmented reality,robotics,image beautification,intelligent agriculture and so on.More and more products need image segmentation algorithm based on deep learning as technical support.Brain tumor is a kind of abnormal tissue caused by uncontrollable factors leading to cell proliferation,which seriously threatens human life and health,and brain tumor is one of the diseases difficult to be conquered in the medical field.Magnetic resonance imaging(MRI)images of brain tumors are segmented to analyze the exact location of areas of edema,enhancement,and necrosis.This plays a key role in the later diagnosis and treatment.In traditional method of brain tumor segmentation,radiologists use specific software to manually segment and label data based on their knowledge of anatomy and pathology.This method is time-consuming and laborious,and the accuracy rate of labeling is different due to individual ability.Besides,this method is unstable.Therefore,the traditional methods of brain tumor segmentation are difficult to meet the clinical needs in terms of segmentation speed and precision.In view of the above problems and background,this paper completed the following works:1.Proposed an image segmentation model based on path aggregation U-net.We applied this model to brain tumor segmentation.The traditional image segmentation model of encoder-decoder has some problems,such as insufficient decoding ability,insufficient feature aggregation and excessive use of computing resources.The path aggregation U-net can enhance the decoding ability of the model,aggregate more shallow features from segmentation results,and reduce the computational resource occupancy rate in comparison with traditional deep learning pyramid model.And the model can improve the segmentation performance.2.Proposed an image segmentation model based on feature mining network.We applied this model to brain tumor segmentation.This model raised the semantic information mining unit(SIMU),macro information mining unit(MIMU)and feature correction unit(FCU).These three units can increase the mining of semantic information and spatial information,and further correct information in a direction conducive to the segmentation result.Each unit can significantly improve segmentation performance.