Brain tumours represent critical medical conditions requiring accurate and timely diagnosis to improve patient outcomes and guide effective treatment strategies. Manual interpretation of magnetic resonance imaging (MRI) scans by radiologists remains time-consuming and subject to inter-observer variability. This study addresses these challenges by proposing an ensemble deep learning framework that integrates three complementary convolutional neural network architectures: ResNet101V2, InceptionV3, and EfficientNetB0. The methodology employs transfer learning from ImageNet pretrained weights, leveraging global average pooling to extract discriminative features from brain MRI scans. The ensemble system classifies images into four categories: glioma tumours, meningioma tumours, pituitary adenomas, and normal brain tissue. Comprehensive experimental evaluation on a dataset of approximately 3,000 MRI images demonstrates an overall classification accuracy of 82 percent, with precision, recall, and F1 Score of 84 percent, 82 percent, and 80 percent, respectively. Class-specific analysis reveals exceptional performance for pituitary tumour detection, with 97 percent precision and 92 percent recall, while meningioma classification achieves 97 percent recall. The ensemble approach outperforms individual architectures by capturing complementary feature representations across multiple scales and hierarchies. These results demonstrate the clinical potential of ensemble deep learning for automated brain tumour diagnosis, offering a robust framework that balances computational efficiency with diagnostic accuracy. The proposed system provides a foundation for future development of clinical decision support tools in neuro-oncology.