Classification of Brain Tumors Using a Hybrid VGG16-ResNet50 Architecture with SVM

Abstract

Brain tumor classification from Magnetic Resonance Imaging (MRI) is critical for clinical diagnosis and treatment planning. This study proposes a hybrid deep learning architecture combining VGG16 and ResNet50 convolutional neural networks, utilized strictly as fixed feature extractors via transfer learning, followed by Principal Component Analysis (PCA) for dimensionality reduction and Support Vector Machine (SVM) classification. Unlike approaches that train CNNs from scratch, we leverage pre-trained ImageNet weights to extract high-level feature representations without fine-tuning the convolutional layers. The methodology was evaluated on a dataset of 3,064 T1-weighted contrast-enhanced MRI images categorized into three tumor classes: Meningioma, Glioma, and Pituitary tumors. Three experimental configurations with distinct train-test ratios (80:20, 70:30, and 60:40) were investigated using strict patient-level splitting to prevent data leakage, achieving 90.36% accuracy (80:20 split) with 0.9564 specificity. To rigorously validate clinical generalization, a 5-fold cross-validation was subsequently performed, yielding 85.90% accuracy. The consistent high specificity (>0.91) across all configurations indicates reliable clinical applicability. The fused feature representation, combined with PCA compression to approximately 290 components and SVM margin maximization, produced a computationally efficient and interpretable model suitable for deployment in resource-constrained clinical environments. Results demonstrate that architectural fusion of complementary deep learning features with classical machine learning classifiers achieves competitive performance with enhanced generalization and data efficiency compared to single-model approaches. The novelty of this study lies in the empirical validation of a "frozen-fusion" strategy that eliminates the need for expensive backpropagation during training while outperforming standard transfer learning baselines, specifically determining the optimal PCA-SVM configuration for maximizing specificity in pituitary tumor diagnosis within resource-constrained environments.