Accurate detection and classification of brain tumors from MRI scans are vital for timely diagnosis and treatment planning. Convolutional Neural Networks (CNNs) have demonstrated significant promise in this domain; however, their performance largely depends on the careful selection of hyperparameters such as learning rate, optimizer, activation function, pooling type, batch size, number of convolutional layers, and epochs. Most existing studies rely on automated optimization techniques like genetic algorithms, grid search, or Bayesian optimization, which operate as black-box approaches with limited interpretability, while others use arbitrary or partially tuned hyperparameters without systematic experimentation. To address these gaps, this research conducts an extensive manual hyperparameter optimization process across seven key parameters through six controlled experiments using a publicly available Kaggle MRI brain tumor dataset comprising 3,000 MRI images of both tumorous and non-tumorous brain classes. The results reveal that the Adam optimizer with a learning rate of 0.003 provides the best balance between convergence stability and classification accuracy. Using this configuration, a customized CNN model named MTBTDet_CNN was developed for MRI-based brain tumor detection and classification. Experimental findings demonstrate that the proposed model achieves superior performance across multiple evaluation metrics - including accuracy (0.997), precision (0.993), recall (1.000), F1-score (0.997), and specificity (0.993), outperforming existing optimization and manual tuning approaches, thereby validating the effectiveness of the proposed manual tuning strategy.