Feature Selection Method On Glioma For Radiomics

Glioma is the most common primary malignant tumor of the central nervous system,which has the characteristics of high incidence,high recurrence rate,high mortality,high disability rate and low cure rate.Accurate diagnosis of glioma before surgery has become an important prerequisite to save the life of patients.Radiomics is an emerging medical image analysis technology,which uses various statistical analysis and data mining methods to achieve the grading prediction of tumors through highthroughput feature extraction of the areas of interest in the images.Since the features calculated by Radiomics are high-dimensional,it is often difficult to obtain a large number of high-quality samples due to the particularity of medical image data sets.If the high-dimensional features are directly used for model training,the phenomenon of overfitting is likely to occur.Therefore,how to select high dimensional features is a major challenge for Radiomics.Aimed at the characteristics of glioma images and considered the influence of tumor boundaries on the characteristics of glioma.This paper proposed two new feature selection algorithms for the Radiomics study of glioma based on the existing feature selection algorithms.(1)A multi-level feature selection Algorithm(MSOM-GA,Mean Score of mixgenetic Algorithm)based on Genetic Algorithm was proposed.Through the analysis of the existing filtering,embedded and wrapper feature selection algorithms,it is found that the single feature selection algorithm cannot fully take into account the feature relativity and redundancy.In addition,different feature selection algorithms have different emphases,and the selected feature subsets are quite different,which resulted in unstable training results.Therefore,in this paper,the intra-group correlation coefficient is firstly used to select the stability.In order to solve the problem of single evaluation criterion,this paper proposes a feature selection algorithm(MSOM)combining F-Score and information gain as correlation feature selection.After that,the genetic algorithm is used for three-level screening to remove redundancy.The results were graded on the glioma data set of Henan provincial people's hospital,and the results showed that the feature selection algorithm improved the accuracy of glioma grading.(2)A multi-level feature selection Algorithm(MSOM-IAGA,Mean Score of mixed-improve Adaptive Genetic Algorithm)based on improved Adaptive Genetic Algorithm was proposed.This method based on algorithm(1),and improve the genetic algorithm.The genetic algorithm adopts fixed crossover probability and mutation probability,which is likely to generate the phenomenon of "precocity".Although the crossover probability and mutation probability can be dynamically change by the adaptive genetic algorithm according to the algebraic of population evolution,it will occur that the individuals with the greatest fitness in the contemporary population do not carry out genetic operation,so that the algorithm may obtain the local optimal solution rather than the optimal feature in the real sense.Therefore,this algorithm improves the adaptive crossover probability and mutation probability,so that individuals with large fitness in the population also have crossover and mutation probability,which prevents the population from falling into the local optimal.The results of the feature selection algorithm were graded on the glioma data set of Henan provincial people's hospital.The results showed that the feature selection algorithm significantly improved the accuracy of glioma grading compared with the proposed algorithm(1).