Hybrid convolutional neuro-fuzzy networks for diagnostics of mri-images of brain tumors

Abstract

The problem of medical images of brain analysis and classification of detected tumors into two classes: benign and malignant is considered. For brain tumors recognition hybrid convolutional; network was developed in which CNN VGG-16 and ResNetV2_50 were used for feature extraction and FNN ANFIS- for classification of detected tumors. For ANFIS training the adaptive stochastic gradient method was suggested and implemented and its efficiency was explored. For preventing overfitting two layers of dropout were added. As a loss-function binary cross-entropy function was used. As the optimization algorithm Adam W with technique Warm-up was used. The experimental investigations of the suggested hybrid CNN-ANFIS network in the problem of classification real images were carried out on the special data set Brain MRI images for brain tumor detection. The comparisons of classification accuracy of the suggested CNN-ANFIS network with results of convolutional networks with classification algorithms SVM, NN and dimensionality reduction were performed which confirmed the reasonableness of application of hybrid networks for medical images recognition. In general the classifier SVM has shown the best results for considered problem.