Accurate diagnosis of brain tumors is necessary for appropriate treatment with the use of magnetic resonance imaging. This paper investigates the performance of several deep learning models that classify brain tumors through different approaches proposed in the literature. Two types of models were explored to classify four subtypes of brain tumors, namely a custom-built MobileNet-BT model and the pre-trained model VGG-16; the importance of transfer learning in medical imaging becomes apparent. A CNN model trained from scratch identifies the 1p/19q codeletion-a genetic marker for gliomas-at an F1-score of 96.37%, which indicates that specialized architectures are important for genetic marker detection. On the other hand, the performance of conventional CNNs, such as VGG16 and ResNet-50, and Transformer models with XAI methods showed that CNNs perform better than Transformers when the datasets are smaller in size, while their enhanced tumor visualization was confirmed by XAI. Collectively, these findings highlight the requirement for model designs tailored for interpretations to enhance accuracy, reliability, and decision-making in diagnosis of brain tumors.