A Pre-study on the Layer Number Effect of Convolutional Neural Networks in Brain Tumor Classification

Convolutional Neural Networks significantly influenced the revolution of Artificial Intelligence and Deep Learning, and it has become a basic model for image classification processes. However, Convolutional Neural Networks can be applied in different architectures and has many other parameters that require several experiments to reach the optimal results in applications. The number of images used, the input size of the images, the number of layers, and their parameters are the main factors that directly affect the success of the models. In this study, seven CNN architectures with different convolutional layers and dense layers were applied to the Brain Tumor Progression dataset. The CNN architectures are designed by gradually decreasing and increasing the layers, and the performance results on the considered dataset have been analyzed using five-fold cross-validation. The results showed that deeper architectures in binary classification tasks could reduce the performance rates up to 7%. It has been observed that models with the lowest number of layers are more successful in sensitivity results. General results demonstrated that networks with two convolutional and fully connected layers produced superior results depending on the filter and neuron number adjustments within their layers. The results might support the researchers to determine the initial architecture in binary classification studies.