Brain Tumor Classification from MRI Images Using Hybrid Deep Learning Approaches: VGG19 with SoftMax and SVM Classifiers

Brain tumor classification from MRI images remains a challenging task due to the complex structure and visual similarity among tumor types. Many existing deep learning models achieve high training accuracy but often suffer from limited generalization on small medical datasets. The VGG19 convolutional neural network for feature extraction and a Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF) kernel are integrated in this study's hybrid deep learning architecture to close this gap. The rationale behind this combination is that VGG19 efficiently captures deep hierarchical image features, while SVM-RBF enhances nonlinear decision boundaries and improves robustness against overfitting. The suggested models were trained and verified using a publicly available brain MRI dataset from Kaggle, which consisted of 7023 contrast-enhanced pictures divided into four classes: glioma, meningioma, pituitary tumor, or no tumor. To guarantee a fair assessment, the data were split into subgroups for testing, validation, and training. According to experimental findings, the VGG19-SVM (RBF) model attained 96.2% validation and 97.8% testing accuracy, whereas the VGG19-Softmax model earned 99% training accuracy and 98.4% validation accuracy. The novelty of this work lies in providing a controlled and systematic comparison between two classification pipelines—VGG19 with Softmax and VGG19 with SVM (RBF)—under identical preprocessing, augmentation, and evaluation settings. Unlike prior studies that report single-model performance, this study analyzes stability, misclassification patterns, and class-wise behavior in a unified experimental framework, offering deeper insight into the conditions under which each model performs best. Images are immediately compressed by the system using AWS Lambda, S3, IAM, CloudWatch, and Docker. Strong stability and steady convergence over time were demonstrated by both models. Comparative analysis confirms that the proposed hybrid architecture provides competitive performance with improved class separation, highlighting its potential for reliable and automated brain tumor diagnosis in clinical settings.