Brain tumors are a severe health problem in the entire world, and appropriate and prompt diagnosis is one of the keys to improved patient outcomes. Traditional diagnostic methods and procedures relying on manual interpretation of the Magnetic Resonance Imaging (MRI) images are often time-intensive and vulnerable to human error and inter-professional variability. Despite the possibilities of deep learning models in the case of medical image analysis, the issue of overfitting and generalization is also present. In this paper, a detailed comparative study of the five most popular models based on the application of pre-trained convolutional neural networks has been provided that is, MobileNet, VGG16, VGG19, ResNet101, and InceptionV3. The brain tumors are classified into models using MRI image data. Sophisticated image pre-processing techniques, including image resizing, image normalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), and data augmentation to enhance the image quality, eliminate noise, and balance the dataset, are our methodology. Performance comparison was presented between these models, considering their training and test accuracy, and recall, precision, F1-score and Cohen Kappa score. A further contribution to the diagnostic performance and robustness was made by Ensemble Learning. The study determined the test accuracy of the proposed ensemble model of 99.52% precision, recall, and F1-score of 0.9952 and a Cohen Kappa of 0.9936. These findings suggest that an ensemble deep learning model and strong preprocessing methods can improve the reliability and accuracy of brain tumor detection significantly in real clinical practice.