A Stacked CNN-BiLSTM-SVM Framework with Spatial Attention for MRI-Based Brain Tumor Detection

Automated diagnostic methods are paramount for assisting neurologists to facilitate early detection of brain tumors. Existing methods often rely on single convolutional neural network (CNN) architectures and fixed feature extraction schemes, which limit generalization and trustworthiness. Additionally, most existing models employ transfer learning solely for feature extraction without architectural modifications, which constrain their ability to optimize features tailored to the specific task. Accordingly, this study proposes a stacking ensemble framework for brain tumor classification using magnetic resonance imaging (MRI) images. The methodology begins with standard data augmentation to enhance image diversity. MRI images are then passed through a set of CNNs for feature extraction. Each CNN is embedded with bidirectional long short term memory layers and spatial attention mechanisms to capture spatial-temporal dependencies and ameliorate discriminative feature learning. Support vector machine classifier is employed to dynamically fuse the extracted features. The proposed model achieved classification accuracy of 98.41% using 5-fold cross-validation. The developed model outperformed the individual base CNNs used within the ensemble and other widely adopted architectures such as AlexNet, DarkNet19, and GoogleNet. The proposed system holds strong potential for aiding radiologists in validating manual MRI screenings for brain tumor detection.