Deep Learning Architectures for Brain Tumor Classification in 2D MRI

Abstract

Brain tumor classification from MRI scans is a clinically critical task, and deep learning approaches have shown strong potential in accurately detecting various types of brain tumors (BTs). This paper presents a comparative evaluation of seven deep learning architectures for BT detection: VGG-16, ResNet-18, ResNet-50, DenseNet-121, EfficientNet-B0, ConvNeXt-Tiny, and ViT-Base/16. The models are assessed using two publicly available datasets: the Kaggle brain tumor dataset (7,023 images across four classes) and the Figshare dataset (3,064 images across three classes). To ensure robust and unbiased evaluation, a leakage-aware experimental protocol is adopted, incorporating an 80/20 development–test set split along with three-fold cross-validation. When trained exclusively on the Figshare dataset, ViT-Base/16 and VGG-16 achieve relatively low accuracies of 56.12% and 46.49%, respectively. However, the application of transfer learning substantially improves their performance, increasing accuracy to 98.21% and 97.88%. On the Kaggle dataset, EfficientNet-B0 achieves the best overall performance, reaching 99.07% accuracy and 99.70% specificity. Furthermore, an analysis of model accuracy versus parameter size demonstrates that DenseNet-121 and EfficientNet-B0 require approximately four and six times fewer parameters than ConvNeXt-Tiny and ResNet-50, while maintaining competitive performance. These findings suggest that compact architectures can provide an effective balance between classification accuracy and computational efficiency.