Brain tumor detection through magnetic resonance imaging (MRI) is a difficult investigation to conduct. Developing a fast and reliable clinical decision making tool is paramount. Modern techniques like deep learning and convolutional neural networks (CNNs) have demonstrated great promise in automating the process of detecting tumor masses from MRI scans. In this study, we train a VGG16-based CNN, merging two publicly available datasets (Figshare and Kaggle), introducing inter-dataset variability simulating real-world diagnostic conditions. We preprocess data, use stratified splitting for training, testing, and validation, and apply data augmentation techniques. Our model achieves a validation accuracy of 84.4% and demonstrates consistent performance across tumor types. Grad-CAM heatmaps highlight tumor regions with reasonable precision, enhancing model transparency and trust. This work highlights the effectiveness of a lightweight, generalizable CNN architecture along with visual interpretability.