This research investigates data-driven approaches for identifying brain tumors in magnetic resonance imaging (MRI) scans. Using deep learning (DL) methodologies, we develop and optimize two advanced convolutional neural network architectures—ResNet 50 and Inception V3—adapted for categorizing brain MRI images. Building on the proven performance of these models in other medical imaging applications, the present study examines their effectiveness across datasets that differ in complexity, including a four-class dataset and a binary-class dataset. The main contribution of this work stems from the systematic assembly and preparation of these paired datasets. To improve training efficiency, we applied key callback mechanisms such as Early Stopping and ReduceLROnPlateau, along with comprehensive hyperparameter tuning. This involved adding layers, testing multiple loss functions and learning rates, and implementing dropout regularization to support reliable model convergence. Moreover, introducing specialized pooling layers and refined regularization techniques markedly boosted classification performance. Notably, pairing ResNet 50 with the Nadam optimizer produced exceptional results, with class-specific accuracies reaching 99.34% for gliomas, 93.52% for meningiomas, 98.68% for non-tumorous cases, and 97.70% for pituitary tumors. These findings demonstrate the strength of our customized framework, yielding an overall test accuracy of 97.68% across the four categories. In the two-class scenario, ResNet 50 combined with the Adam optimizer achieved superior precision, recall, and F1 scores, along with an overall accuracy of 99.84%. It further delivered per-class accuracies of 99.62% for ‘Tumor Positive’ and 100% for ‘Tumor Negative’, marking a substantial improvement in automated brain tumor detection. This study highlights the potential of optimized DL architectures and targeted training strategies to advance non-invasive brain cancer diagnosis using MRI data.
