Efficient Brain Tumor Classification Using Filter-Based Deep Feature Selection Methodology

Neuroimaging plays an elemental role in disease detection in the domain of medical science. Brain magnetic resonance imaging (MRI) helps to detect chronic diseases such as brain tumors, strokes, and dementia. It is a nonintrusive and sensitive method for evaluating brain tumors. Numerous deep learning techniques have been proposed to analyze brain tumors as they revolutionize feature selection by automatically extracting relevant features, outperforming traditional methods in accuracy and speed. Our paper proposes a first-of-its-kind, two-stage framework for classifying brain tumors from structural MRI scans. In the first stage, a pre-trained convolutional neural network has been used to extract relevant features, considerably reducing training time and the need for extensive hardware. Next, a filter-based deep feature selection method narrows down the high-dimensional features obtained from the previous stage, minimizing computational load and overfitting risks. Finally, a polynomial-kernel Support vector machine performs multi-class classification. We have also employed fivefold cross-validation to ensure reliable results that are not overly sensitive to specific training or testing data. On the first dataset, this paper achieved 98.17% classification accuracy, 97.92% precision, 97.95% recall, and an F1-score of 97.92% while simultaneously reducing 25% of the extracted features. The approach has also been tested on two additional brain tumor datasets, giving classification accuracies of 99.46% and 98.70%. These promising results underscore the potential of our lightweight framework’s robust nature and generalization capabilities, making it suitable for deployment in real-time environments with limited technological resources.