Limited datasets and model underfitting or overfitting are especially common issues in the analysis of medical images. Even while convolutional neural networks can outperform more conventional machine learning-based techniques in terms of results, there are still many difficult problems to be solved, including enhancing their generalization capabilities and creating reliable models. In this research, we offer a framework utilizing transfer learning and enhancing the categorization of brain cancers by data augmentation in Magnetic Resonance Imaging images. Leveraging three pretrained CNNs: ResNet-50, VGG16, and ResNet-101. The purpose of this work is to identify brain tumors using a dataset sourced from Kaggle. Techniques employed for augmenting the data such as shear, zoom, and flip were enhancing the resilience and efficiency of the models. The productiveness of data augmentation was evaluated by comparing model performance with and without augmentation across key metrics: accuracy, recall, and precision. Results indicate that data augmentation led to varying degrees of performance improvement in all models. Notably, ResNet-50 and ResNet-101 saw significant gains in recall, enhancing their sensitivity and ability to detect relevant instances, while VGG16 demonstrated high performance across all metrics in both scenarios. Among the models, VGG16 achieved favorable results with an accuracy of 93%, a recall of 94.67%, and a precision of 91.61%, surpassing ResNet-50 and ResNet-101. This analysis highlights the potential of combining transfer learning (TL) and data augmentation for automated and reliable brain tumor classification.