In this paper, a comparative study of two neural network architectures: firstly, a custom-designed Convolutional Neural Network (CNN) and secondly, the prominent VGG16, for the brain tumor classification within MRI images is presented. A detailed examination of neural network architectures, dataset characteristics and generalization capabilities guiding comprehensive performance is reported. In the custom-designed CNN architecture, three convolutional layers with varying filter sizes, max-pooling layer accompanied by densely connected layers and finally softmax activation function generating the resultant output. Whereas, the VGG16 architecture includes a deeper layer that includes multiple convolutional blocks incorporating 3 × 3 filters each followed by max-pooling layers. Using 20 training epochs, the study systematically evaluates both the neural network models with respect to parameters accuracy, precision, recall, and F1 score. The custom-designed CNN architecture demonstrated a remarkable training accuracy of 99.72%, even if with a noticeable performance gap in the validation set of 73.86%. However, VGG16 architecture demonstrated a training accuracy of 93.31%, even with more better validation accuracy of 76.14%. The study clearly implies the selection of appropriate neural network architecture through its advantages for accurate brain tumor classification in MRI images.