Brain tumor detection is one of the most critical processes in medical imaging, and magnetic resonance imaging (MRI) plays an important role in the diagnostic process thanks to its high-resolution soft tissue details. However, manual evaluation of MRI images is atime-consuming and error-prone process. Therefore, the development of deep learning-based automated systems is of great importance in terms of increasing diagnostic speed and improving the effectiveness of healthcare services. In this study, the aim was to classify brain tumors using a dataset consisting of Brain MRI images. First, image processing techniques were applied to the dataset to obtain enhanced images. Then, the original and enhanced images were fed into seven different transfer learning architectures. The results showed that enhanced images provided higher accuracy in all models. In the second phase of the study, different activation functions were tested on the two most successful models, ResNet101 and ResNet152. With the proposed hybrid activation function, ResNet101 achieved 98.42% accuracy and ResNet152 achieved 97.20% accuracy, demonstrating the highest performance. These findings reveal that the proposed method significantly improves classification success.