AN INTEGRATED APPROACH FOR BRAIN TUMOR DETECTION USING DEEP FEATURE EXTRACTION AND MACHINE LEARNING CLASSIFIERS

Abstract

Brain tumor classification using medical image processing is becoming easier with the fast pace of innovation in medical image analysis technology which has a large amount of MRI data sets related to brain tumors. The precise classification of brain tumors is vital for diagnosing and treating brain disorders. The current research project aims to design a hybrid computer-aided brain tumor classification system based on deep learning algorithms and conventional methods. In contrast to the previous approach where only end-to-end deep neural networks have been used, this paper uses the DenseNet201 architecture as a pretrained architecture to learn high-level deep visual features from the MRI images. The learned deep features are further refined by applying the Extra Trees Classifier based feature selection process in order to retain only the most valuable features and reduce the redundancy. Finally, different supervised classifiers like Logistic Regression and other machine learning classifiers are used for evaluation purposes. The experiments have been done using a publicly accessible Brain Tumor MRI Dataset provided by Kaggle, which consists of four types of The following classes were considered in this study: glioma, pituitary tumor, meningioma, and non-tumor. The criteria adopted to measure the efficiency of the proposed technique included accuracy, precision, recall, and F1 score. It was seen that both the models Linear SVM and Logistic Regression performed well during classification. This demonstrates that there is a way of coming up with tumor classification systems that are both efficient and accurate, and yet do not require a lot of computing resources, unlike in the full training of deep neural networks. The system will be very instrumental in aiding medical professionals in diagnosing brain tumors. In addition, there is enough room for development, such as explainable artificial intelligence and attention maps for tumor localization.