Research On Segmentation And Grading Of Glioma MRI Images Based On Multi-Scale Features

Glioma is a common primary brain malignancy with high morbidity and mortality.The timely detection of brain tumors in clinical practice and the development of corresponding treatment plans can improve the cure rate and survival rate of patients.Magnetic Resonance Imaging(MRI)is non-invasive and has a good imaging effect on brain tissue,which can help doctors differentiate brain tumor areas from normal tissue.At present,doctors manually segment various regions of glioma in MRI images.However,due to the different working experiences of doctors,the results of tumor region determination are different,and the segmentation results are less repeatable and more subjective.Therefore,the realization of automatic segmentation of brain tumors has important clinical significance.With the development of machine learning and deep learning,people’s research on MRI images has gradually developed from artificially designed models to construct features to automatically extract features from deep learning models.The scale is no longer limited to shallow image features but through convolution extend to deep features.Based on this,this paper takes MRI images of glioma as the research object and proposes a set of segmentation and grading methods for MRI images of gliomas based on multi-scale features,which can effectively assist clinical accurate diagnosis and treatment.The specific content of this article is as follows:In the aspect of shallow feature extraction model construction,because of the problems of low sample delineation efficiency and strong subjectivity of multi-sequence MRI images,a glioma segmentation method based on multi-sequence feature construction is proposed.By fusing the spatial domain and frequency domain information of the image,the feature set used to represent the structural information of brain tissue is extracted,which solves the limitation of traditional single domain feature extraction.Based on the idea of graph semi-supervised learning,a local-global adaptive information learning algorithm is proposed to complete the segmentation of gliomas.The experimental results show that the method can ensure tumor segmentation accuracy and improve the segmentation efficiency under the premise of less labeled samples,and the generated quantitative indicators can provide the basis for the early clinical diagnosis of glioma MRI.While the good segmentation results by that method,the segmentation algorithms based on shallow features still have certain limitations in the design of feature extraction models.The commonly used convolutional neural network can obtain deep features,which will cause the problem of local information loss.To enhance the algorithm’s ability to perceive local related information,a graph convolutional neural network segmentation method based on dilated neighborhoods is proposed.The edge connection mechanism of the dilated neighborhood is proposed to expand the image data of U-Net into graphstructured data,and solve the problem of edge construction in the transformation of image data into graph-structured data.Then,GCN is used to classify the graph structure data to realize the segmentation of glioma regions.The experimental results show that compared with U-Net,this method has obvious improvement in various indicators.Clinical MRI image analysis of glioma includes lesion segmentation and grading diagnosis.Quantitative analysis of the segmented lesion area helps to improve the accuracy of automatic grading,thereby providing a reliable basis for auxiliary diagnosis.To this end,combining shallow features with deep learning features,an automatic grading method for gliomas based on multi-scale features is proposed.Therefore,utilizing deep features expands the ability to acquire features and improves the performance of model predictions.The experimental results show that this method can accurately segment and grade brain gliomas,which is helpful for clinicians to make accurate diagnoses and formulation of treatment plans.