Brain Tumor Grading Based On Conventional Magnetic Resonance Images

The method of brain tumor grading is currently based on histopathologic analysis which limits clinical application because of kinds of disadvantages such as sampling error or invasion.A method based on conventional magnetic resonance images was proposed to grade brain tumor.Kinds of features were extracted from region of interest including shape features,intensity features and texture features and they were selected by hybrid feature selection methods to distinguish the HGG and the LGG.Three classification algorithms based on feature selection method of filter were compared: support vector machine,decision tree and K nearest neighborhood? The result showed that supported vector machine has the best performance of accuracy and sensitivity which was applicable for brain tumor grading.The wrapper method using genetic algorithm was compared with filter method,and the result showed that specified combination of features showed better performance.We need to extract the features after segmentation of tumor in the application of brain tumor grading.Thus,the study of segmentaion of brain tumor is of great importance.An improved method based on the level set algorithm was proposed to segmentate the FLAIR images contained brain edema,which used Otsu algorithm to preliminarily locate the edema area and provide initial contours.We compared two kinds of level set algorithm,one of which is edge-based,the other is region-based.We selected slices from the BRATS2017 database which were different in tumor homogeneity and edge to carry out our experiments.The result shows that the region-based algorithm is better than that of the edged-based.Subjectively,the contour derived by the region-based algorithm is closer to the real one.Objectively,all evaluation indexes besides precision show that the region-based algorithm is better.