This research aims to address an urgent necessity of correct and timely identification of brain cancer in MRI images at early stage by using deep learning models for improving diagnostic accuracy as well help plan the treatment. The increasing number of brain cancer occurrences and the diagnostic challenges in this area necessitate specific, automatic systems to help clinicians. This study uses a dataset which includes 2800 MRI images in different region scores and stage having gone through preprocessing steps followed by feature extraction then from these trained on three state-of-the-art models: DenseNet169, MobileNetV3 and Resnet152. The results show that DenseNet 169 gives the highest accuracy of 97.68% along with excellent precision, recall and F1 score moreover implying its better aptitude in finding minute changes in cancer stages. MobileNetV3, on the other hand offers a sweet spot of 94.5% accuracy and provides good trade-off between compute efficiency, i.e., computational cost is less, suitable for real-time applications like AI auto focus. Even the more complex ResNet 152 records an accuracy of just 91.23% with higher error rates, suggesting areas for deep network architecture to be explored. The performance analysis (including confusion matrices) outlines the robustness of DenseNet 169 and practical feasibility of MobileNetV3, whereas it points out limitations ResNet 152. This complete assessment not only propels the medical imaging discipline forward, but it also additionally gives vital lessons for better diagnostic tools.