Comparative Analysis of Modern CNN, Transformer, and Lightweight Architectures for Multi-Class Brain Tumor Classification from MRI Data

Early diagnosis of brain tumors is vital for clinical treatment success. This study comparatively analyzes the performance of Xception, MobileNetV3-Large, ConvNeXt-Tiny, and Vision Transformer models in classifying magnetic resonance imaging data in four different categories (glioma, meningioma, pituitary tumor, and no tumor). Experimental findings revealed that under the provided conditions, the Xception architecture exhibited the highest accuracy, offering superior stability, particularly in distinguishing healthy tissues. While the MobileNetV3-Large and ConvNeXt-Tiny models yielded competitive results, the ViT model showed limited success compared to the other models. In terms of computational efficiency, MobileNetV3-Large was the fastest, while ConvNeXt-Tiny was the most expensive model. The results demonstrate that powerful convolutional neural networks offer a strategic advantage in medical decision support systems requiring high accuracy.