Cancer remains the second leading cause of death globally, with brain tumors exhibiting some of the lowest survival rates among all cancer types. Accurate diagnosis, guided by the tumor’s structure and location, is essential for selecting appropriate treatment strategies and improving patient outcomes. This study proposes a novel deep learning approach for classifying brain tumors from magnetic resonance imaging scans, aimed at enhancing diagnostic precision. Given the growing reliance on computer-aided diagnosis (CAD) systems, there is a pressing need for tools that can assist radiologists in detecting and categorizing brain tumors more effectively. We conducted a comprehensive evaluation of several pre-trained deep learning models across three distinct datasets to determine the most effective architecture for brain tumor detection. Based on this analysis, we developed BrainNet, a custom convolutional neural network (CNN) optimized for MRI-based tumor classification. BrainNet employs multiple layers of convolution and pooling, followed by dense layers to extract and learn discriminative features. The model categorizes brain tumors into four classes: Meningioma, Glioma, Pituitary, and No Tumor, using a softmax output layer. Despite leveraging transfer learning techniques, BrainNet consistently outperformed well-established pre-trained models, demonstrating superior accuracy, precision, and efficiency. Our experiments across multiple datasets confirm that BrainNet achieves a classification accuracy of 99.92%, along with excellent recall and F1-scores. Its lightweight design and high accuracy make it a promising solution for deployment in real-world clinical environments, including resource constrained settings.
Khan et al. (Tue,) studied this question.