This study proposes a deep learning-based approach to classifying brain tumors using MRI images to enhance accuracy and stability for different kinds of tumors. Three new datasets from Mendeley Data are used: D1 (PMRAM: Bangladeshi Brain Cancer – MRI Dataset), D2 (Brain Tumor MRI Dataset with four types: Meningioma, Pituitary, Glioma, and No Tumor), and D3 (Brain Tumor Data). For training and testing purposes, D1 and D2 were combined to form a dataset of 13,664 MRI images (Meningioma: 3209, Glioma: 4173, Pituitary: 3450, No Tumor: 2832). D3 was used as an external validation set of 7023 MRI images (Meningioma: 1645, Glioma: 1621, Pituitary: 1757, No Tumor: 2000). A few of the state-of-the-art architectures—VGG19, InceptionV3, Xception, MobileNetV2, ResNet50V2, and DenseNet201—were compared against the proposed hybrid model, BrainNet. BrainNet achieved the highest model accuracy of 99% on the combined dataset and provided exceptionally well performance on the external validation set with an overall accuracy of 98.18%. It recorded the least error values among all models with MSE 0.0184, RMSE 0.1358, and MAE 0.0128, while other models achieved VGG19: 93.12%, InceptionV3: 97.16%, Xception: 97.41%, MobileNetV2: 97.87%, ResNet50V2: 98%, and DenseNet201: 98.22%. These results confirm the efficacy and accuracy of BrainNet in computerized brain tumor diagnosis and its feasibility in the clinical environment.