Research And Implementation Of Detection Algorithm Based On Multi-modal MRI Image Of Glioma

Glioma is the most common primary brain tumor with high morbidity and mortality.Even with the most aggressive treatment,the prognosis is poor.The methylation status of the molecular biomarker O6-methylguanine-DNAmethyltransferase(MGMT)promoter is closely related to the chemotherapeutic effect and prognosis of glioma patients,so the detection of its status is of great significance to patients.Most current detection methods are based on invasive detection of living tissues.In this paper,a quantitative image analysis method is proposed to predict the methylation status of MGMT promoter.The method mainly includes image segmentation,feature extraction and classification prediction.This paper will focus on these three aspects and carry out specific research.The segmentation of image is the key of the whole process,and the results directly affect the subsequent analysis.In this paper,a fully automatic segmentation method based on multimodal convolutional neural network is proposed to segment patient MRI images into five subregions: edema,enhancing,non-enhancing,necrosis and health.In order to make full use of the difference information of multimodal images,a convolution neural network feature extractor is constructed for each modal image,and then combine the feature to train the model,thus to optimize the segmentation results.Then,two convolution layers with small convolution kernel are used in the model to replace the convolution layer with a large convolution kernel.we use the small convolution in the convolution layer to construct a deeper model,so we can improve the discriminant ability of the model.The validation results of the public dataset show that this method can effectively and accurately segment the subregions of glioma.Compared with other algorithms,this method can segment the tumor more comprehensively.Feature extraction is the core of the whole quantitative image analysis,the histogram features and texture features are widely used in image analysis.This paper combines multimodal images and multi-region of interest,using histogram analysis method and gray-level co-occurrence matrice,gray-level run length matrice,gray-level size zone matrice,neighborhood gray-tone difference matrices statistical texture analysis methods to obtain a large number of first-order,second-order and higher-order features,so as to provide more powerful support for the relationship between the image features and MGMT methylation analysis.For the extracted image features,the correlation feature selection method is used to remove the irrelevant and redundant features,and to provide an imaging signature with significant discriminatory power for MGMT methylation status detection.In order to predict the methylation status of MGMT promoters,two classification algorithms,Random Forest(RF)and Support Vector Machine(SVM),are used to construct the prediction model,the leave-one-out cross validation is used to train and validate model.Through experiments,the AUC values based on the RF and SVM models are 0.898 and 0.768 respectively.The experimental results show that the quantitative image analysis model designed in this paper is effective and accurate in the detection of MGMT methylation status,and the RF model has better performance.Compared to the existing MGMT state detection methods for quantitative image analysis,this paper fully considers the combination of potential features of multiple imaging modalities,regions of interest,and texture feature analysis methods to construct a detection model.