<p>Brain tumors are abnormal tissue that may become cancerous and harm a person. Brain MRIs show them. These methods rely heavily on segmentation and classification. Tumors vary in size, location, and form, making tumor segmentation harder than anatomical organ segmentation. Thus, deep learning-based approaches must be dependable, accurate, and effective. Recent deep learning brain tumor classification and segmentation methods showed promise. These techniques have hefty architectures and need additional storage and expensive training processes due to the large number of training parameters they must feed. Investigation of transportable deep neural network models without reducing classification accuracy is critical. We create compact deep neural network models utilizing pre-trained Attentiveness MobileNetV2 models and the attention module in this study. The suggested system comprises four phases: preliminary processing, division, feature extraction and categorization, and severity classification. First, anisotropic diffusion processing and data improvement are applied. The modified dimensional U-Net (3D-M-U-Net) segments the tumor. Compact MobileNetV2 implements feature extraction and categorization. High-level tumor-based information is first extracted from convolution features. Attention modules gather semantically significant data. After combining high-level tumor-based data with exciting semantic information in the convolutional and focused modules, fully connected layers and the softmax layer classify tumors as benign or harmful. Finally, SVM classifies tumors as moderate, severe, or mild. The proposed method was evaluated using Brats-2020 and Brats-2019 high-quality brain cancer pictures. The proposed model outperforms standard and hybrid models in precision, recall, accuracy, F-Score, DSC, and SSIM. It was the most productive and successful strategy tried. The recommended model has 99.9% accuracy, precision, and recall across both datasets.</p>