The design of neural network architectures often relies on heuristic decisions. In this study, we systematically varied the number of convolutional filters, convolutional blocks, and dense-layer widths in convolutional neural networks (CNNs) and evaluated their performance on a brain tumor classification task. Using 682 model configurations and repeated training runs, we observed that classification accuracy depends strongly on the chosen architectural design. Multiple network configurations achieved competitive performance, indicating that efficient models can be identified without relying on large architectures. These findings highlight the importance of structured architecture exploration for constructing effective and resource-efficient neural networks in biomedical imaging.
Bindemann et al. (Mon,) studied this question.