Multi-modal Brain Tumor Image Segmentation Based On Mixed Attention Mechanism

In recent years,with the continuous decline of air and environmental quality,as well as people’s inability to control the intake of nitroso compounds in their daily diet,brain tumor has become one of the common tumors threatening people’s health.With the development of medical imaging,magnetic resonance imaging(MRI),as one of the imaging technologies,plays an important role in doctors’ diagnosis process.However,doctors’ judgment of images is mostly based on subjective experience,and different doctors have slightly different judgment of images.This wastes a lot of energy on image discrimination,which delays the timely diagnosis of patients.Image segmentation,as a hot research topic in the field of computer vision,has been widely used in the field of brain tumor image segmentation to assist doctors in rapid diagnosis.However,there are some problems in brain tumor images,such as high noise,unclear texture features in the lesion area and insufficient linear features in the image.And,as a medical image itself,it involves the privacy of patients.Professional and experienced radiologists are required to annotate the images so that the results can be more convincing.In recent years,many domestic and foreign researchers have made great progress in the field of medical image segmentation,especially brain tumor segmentation,especially the image segmentation based on deep learning technology,which has excellent performance in many organ lesion segmentation tasks.However,there are still some problems in most of the methods.For example,the deep segmentation model is not enough to mine the high-dimensional features of 3D images,the boundary segmentation between the focal area and normal organs is fuzzy,the model itself has many parameters,and the floating point calculation of the model is large,which makes the reasoning speed of the model slow and the generalization ability is low.Based on the above problems,we proposed a 3D brain tumor image segmentation architecture based on Hybrid Attention mechanism called Hybrid Attention Network.HANet adopts a decoding method,including 3D feature encoder decoder and 3D characteristics,and in the output stage of the decoder,we introduced based on channel attention mechanism and mechanism of spatial attention attention mechanism module,and put forward a kind of attention AGUs door unit,used to integrate the encoder on the stage and sampling process characteristics of two parts.The network adopts the way of dense connection to prevent the loss of parameters in the network training process.The main research contents of this paper are as follows:Firstly,the background and research significance of brain tumor medical image segmentation are introduced,and the research status of image segmentation in medical images,especially brain tumor image segmentation is summarized,and the advantages and disadvantages of this method are analyzed.This paper introduces the characteristics of medical image of brain tumor,and briefly introduces the correlative techniques of brain tumor image segmentation.Secondly,a new segmentation architecture for brain tumor images,HANet,is proposed.The architecture adopts a V-shaped structure,adopts codec mode,and adds mixed attention mechanism and attention gate unit.The architecture also adopts the way of intensive connection,which prevents data loss on the one hand,and improves the speed of the network on the other hand.Finally,based on the existing data set of Brats18 brain tumor images,we enhanced the data set,expanded the data set,put forward a whole set of preprocessing methods of brain tumor images,and proposed an overlapping image segmentation method.The validity of this method is proved by experiments,and the validity of HANet segmentation architecture is verified under this data set.