Brain Tumor Diagnosis using Hybrid Machine Learning Methods

Abstract

Brain tumor is a life threatening disease, accurate and fast detection will save human life from death. The research is intended to develop an automated framework for detecting benign and malignant brain tumor from the input brain MRI images containing multiple sequences with improved accuracy and reduced computation time employing hybrid machine learning algorithms at various processing stages. newlineA 3D CNN were implemented. The brain tumor MR images consists of four modalities of multiple sequences. Since all modalities carries important information for segmentation, all image sequences of multiple modalities are concatenated spatially and image averaging was done to get the final sequence integrated input image. To reduce the computation overhead of the CNN the brain area from the input MR brain tumor image was separated using grabcut algorithm. The ROI is applied at the input of 3D CNN and the deep features from the last fully connected layer of the CNN was collected for further processing. The feature vectors were applied to the input of the machine learning classifiers SVM, KNN, RF, NB and softmax layer to analyze the impact of deep features on the classification accuracy of individual classifiers which was the second objective. After evaluation of all these classifiers, the SVM provided better accuracy. newlineThe second work implemented GLCM and LDP feature extractors to capture statistical and texture features from the input MRI. To extract deep features, it used ResNet. The sequence integrated image newlinevi newlineis applied at the input of the three feature extractors and the resultant feature vectors were combined spatially. The hybrid feature vector was applied at the input of machine learning classifiers SVM, RF and NB. Since the variability of a single model reduces prediction accuracy, the result of these classifiers are applied to a majority voting module. The output of the module is taken as final class of the current input.