Brain Tumor Segmentation Based On Deep Learning

Brain tumors are abnormal cells that grow in the brain.The source of abnormal cells can be divided into two categories: primary brain tumors and secondary brain tumors.Primary brain tumors are those that arise directly from abnormally growing cells in the brain tissue,and secondary brain tumors are abnormal cells that spread into the brain from normal cells in other parts of the body that become cancerous cells.Gliomas are common primary brain tumors in adults that arise from glial cells and infiltrate surrounding tissues.Despite considerable progress in glioma research,patient diagnosis remains poor.Accurate brain tumor segmentation is a prerequisite for diagnosis and treatment.The main challenge in this task is the appearance,location,and location of gliomas and their subregions.and shape diversity;fuzzy boundaries,with a high degree of grey similarity to adjacent normal brain tissue;complex boundaries between subregions.Therefore,the accurate segmentation of brain tumors in brain MRI has become one of the difficulties and priorities in the field of medical images.Brain tumor segmentation methods mainly include manual segmentation,semi-automatic segmentation,and fully automatic segmentation.Manual segmentation usually requires several senior physicians to spend a lot of time completing the segmentation,which is tedious,timeconsuming,and error-prone.Although the semi-automatic segmentation method of brain tumors can obtain better results than manual segmentation,there are also different segmentation results of different experts or the same user at different times.For fully automated brain tumor segmentation,the computer precisely segments the brain tumor without any human intervention.In general,fully automatic segmentation algorithms combine artificial intelligence and prior knowledge.With the development of machine learning algorithms that can simulate human intelligence for effective learning,research on fully automated brain tumor segmentation has become a hot research topic.At present,the continuous development of deep learning has gradually become the research focus of automatic brain tumor segmentation.This paper mainly studies brain tumor segmentation through deep learning methods.The main contents are as follow:(1)A brain tumor segmentation algorithm based on 3D depthwise separable convolution and spatial and channel squeeze & excitation module is proposed.The 3D depth separable convolution is added,and the 3D depth separable convolution reduces the number of parameters in the convolution operation process by splitting.At the same time,the spatial and channel squeeze & excitation module is introduced to realize the attention mechanism,and the attention mechanism is introduced to explore the independent encoding of spatial and channel modes.This algorithm mainly solves the problem that the algorithm model in the field of brain tumor segmentation generally requires high computing power and cannot achieve fast segmentation of the model.(2)A brain tumor segmentation algorithm based on hierarchical view ensemble convolution is proposed.First,modify 3D UNet is proposed,which has higher advantages in parameter quantity and segmentation accuracy than 3D UNet.In the input data stage,and overall decomposed convolution is used,which reduces the data size for subsequence processing while making all the information available regardless of down-shuffling operation.Secondly,drawing on the previously studied depthwise separable convolution splitting method,a hierarchical view-integrated convolution is proposed,which enables multi-scale and longrange multi-view context information to be encoded in a single module,significantly reducing the amounts of parameters.The algorithm can realize the lightweight of the algorithm of brain tumors,and the segmentation accuracy has been improved.(3)A brain tumor segmentation algorithm based on multi-view hierarchical segmentation is proposed.Multi-view fusion convolution and multi-view hierarchical segmentation modules are proposed,which extract features of image information from axial,sagittal,and coronal views,respectively.The multi-view hierarchical segmentation module combines multi-view fusion convolution and combines the principle of hierarchical decoupling.By increasing the information flow between different features through splitting and joining operations,each group learns different features freely,reducing feature information redundancy and increasing feature learning for blocks.By adding the above modules,the algorithm proposed in this chapter has the characteristics of simple structure,excellent performance,and lightweight,and the segmentation accuracy has achieved the most advanced effect in some indicators.