Customized CNN for Multi-Class Classification of Brain Tumor Based on MRI Images

In this paper, we propose a new strategy to exploit the advantages of Deep Neural Network-based architectures for brain tumor classification using MRI images for a better diagnosis. This was achieved by analyzing and evaluating pre-trained models on three different datasets. To better design the optimal architecture for solving the classification of brain tumor using MRIs, we have conducted extensive experiment-based analysis on how different layers of Convolutional Neural Network (CNN) process the inputs. Four distinct architectures are then built, each with its specific hyperparameters and layers. The images are fed into the convolutional layers for feature extraction followed by a softmax function before applying the classification process. An extensive experimental study carried out clearly demonstrates that our novel CNN-based classification approach achieves state-of-the-art accuracy, precision, recall and an F1-score of 99.76% 99.64% 99.62% and 99.64%, respectively. Also, a higher performance in terms of Micro-Avg Matthew correlation coefficient (MCC) of 0.929 is achieved. This exceptional performance is achieved thanks to the new proposed model's architecture. Indeed, unlike conventional methods, that often rely on complex transfer learning models or hybrid architectures, our approach utilizes a custom and non-hybrid scheme. Consequently, this streamlined architecture offers a significant advantage of being remarkably lightweight, enabling efficient operation on resource-constrained computing systems.