Distinguishing between benign and malignant bladder lesions is critical for determining appropriate treatment plans and improving patient outcomes. Magnetic Resonance Imaging (MRI) provides detailed imaging that has been leveraged in this study. The deep learning approaches have shown promising performance by achieving higher scores in various performance metrics. Also, the ensemble technique outperformed the individual DNN models, indicating its potential to be the best approach for the diagnosis. We evaluate three state-of-the-art DNN models, VGG19, ResNet50, and MobileNetV2 using a publicly available MRI dataset of bladder lesions. Performance metrics include accuracy, positive predictive value, sensitivity, and F1-score. An ensemble method using a max voting classifier combines the strengths of individual models. The ensemble model achieves the highest accuracy of 98.27%, outperforming individual models. It can be said that deep learning approaches hold promise for improving the differential diagnosis of bladder lesions from MRI scans. While practical implications for clinical decision-making have not been implemented yet, this study provides a foundation for future applications in medical practice. Our study contributes by demonstrating the potential of deep learning and ensemble techniques in this context.