Classification of brain MRI using hyper column technique with convolutional neural network and feature selection method

A proper and certain brain tumor MRI classification has a significant role in current clinical diagnosis, decision making as well as managing the treatment programs. In clinical practice, the examination is performed visually by the specialists, this is a labor-intensive and error-prone process. Therefore, the computer-based systems are in demand so as to carry out objectively this process. In the traditional machine learning approaches, the low-level and high-level handcrafted features used to describe the brain tumor MRI are extracted and classified to overcome the mentioned drawbacks. Considering the recent advances in deep learning, we propose a novel convolutional neural network (CNN) model that is combined with the hypercolumn technique, pretrained AlexNet and VGG-16 networks, recursive feature elimination (RFE), and support vector machine (SVM) in this study. One of the great advantages of the proposed model is that with the help of the hypercolumn technique, it can keep the local discriminative features, which are extracted from the layers located at the different levels of the deep architectures. In addition, the proposed model exploits the generalization abilities of both AlexNet and VGG-16 networks by fusing the deep features achieved from the last fully-connected layers of the networks. Furthermore, the discriminative capacity of the proposed model is enhanced using RFE and thus the most effective deep features are revealed. As a result, the proposed model yielded an accuracy of 96.77% without using any handcrafted feature engine. A fully automated consistent and effective diagnostic model is ensured for the brain tumor MRI classification. Consequently, the proposed model can contribute to realizing a more objective evaluation in the clinics, supporting the decision-making process of the experts, and reducing misdiagnosis rates.