A Research Of Multi-Modal MRI-Based Brain Tumor Segmentation Methods

At present,malignant tumors are one of the main causes of death in China.The incidence of brain tumors increases in the past few years,and brain gliomas are the most common in primary brain tumors.With the continuous development of modern imaging technology,especially the MRI,it is helpful to evaluate the condition of brain tumor and select the treatment methods.The accuracy of brain tumor segmentation is the key for doctors to diagnose the patient's disease information.Brain tumors have complex structures,varying shapes,uneven gray levels,and considerable differences in different patients.Manual segmentation of MR brain tumor images is time-consuming and labor-intensive,and is often subject to subjective differences.For the above reasons,many researchers have devoted themselves to the development of semi-automatic or automatic brain tumor segmentation methods.Firstly,this paper uses a segmentation method based on superpixels.SLIC is used for image segmentation after image preprocessing,and then grayscale statistical features and texture features are extracted from superpixel blocks.The grayscale statistical features include mean,standard deviation,skewness and kurtosis,etc.GLCM and LBP are used to extract texture features.Finally,SVM classifier is used to classify each super-pixel block as a tumor area or a non-tumor area.Secondly,this paper uses a U-Net automatic brain tumor segmentation method based on full convolutional neural network.We first used the Elastic transform for data enhancements.After initializing the parameters of U-Net,we use the Adam optimizer for optimization.By constructing a complete U-Net network with optimized parameters,automatic brain tumor segmentation can be achieved.Finally,this paper used DSC?Sensitivity?Specificity as a unified evaluation criteria to compare the two methods.Although the U-net and the superpixels have all achieved satisfactory results,the results of U-Net are excellent and the U-Net is advanced.The two methods provided in this article can generate specific brain tumor segmentation models for patients without artificial intervention,and can have clinical guiding significance for diagnosis and treatment.The two methods provided in this paper can produce specific segmentation for images of brain tumors without human intervention,which is of clinical significance for diagnosis and treatment.