Comparative Analysis of Standard Deep Learning Architectures for MRI Brain Tumor Classification

Magnetic resonance imaging (MRI) serves as the primary diagnostic tool for brain tumors due to its high accuracy. However, manual interpretation can be timeconsuming and susceptible to human error. Automated computer-aided systems help address these limitations, and recent advances in deep learning offer promising support for diagnostic workflows. In this study, the effectiveness of a wide range of standard deep learning models for brain tumor classification is evaluated. After preprocessing, MRI scans undergo feature extraction using twelve benchmark networks including VGG variants (VGG16, VGG19), ResNet architectures (ResNet50, ResNet101, ResNet152), Inception models (InceptionV3, InceptionResNetV2), DenseNet configurations (DenseNet121, DenseNet169, DenseNet201), and MobileNet versions (MobileNetV1, MobileNetV2). The obtained features are then classified using either an SVM or random forest algorithm. We evaluate system performance across multiple metrics to identify the most effective architectures. Results demonstrate that standard deep learning models outperform custom-built networks while offering room for further improvement.