A lightweight multi-path convolutional neural network architecture using optimal features selection for multiclass classification of brain tumor using magnetic resonance images

Abstract

Brain tumor diagnosis requires precision due to the high mortality rate associated with the growth of abnormal cells in the brain. Early disease identification, improved survival rates, and less reliance on professional MRI analysis is possible with computer-aided diagnostic (CAD) systems using advanced technology such as Convolutional Neural Networks (CNN), which have successfully detected brain tumors in MRI images. However, studies often overlook the impact of Rician noise and scaling on deep learning performance, considering varying tumor sizes, locations, shapes, and boundaries. Pre-trained models like AlexNet, Residual Networks (ResNet), and Inception V3 are effective but has high computational costs due to trainable parameters. Therefore, a lightweight Multi -path Convolutional Neural Network (M-CNN) is introduced to extract features using varying convolutional filters at each convolutional layer. Furthermore, an optimal features module implemented to select the most promising features to enhance our proposed multi-path architecture's performance and computational efficiency. In addition, we evaluate the performance of the proposed M-CNN with the Convolutional Neural Network (CNN), Deep Convolutional Neural Network (D-CNN), and other state-of-the-art deep learning architectures using all and selected features for brain tumor detection. The proposed M-CNN outperforms traditional Convolutional Neural Network (CNN) architectures, achieving 92.25% accuracy with all features and 96.03% with selected features. It demonstrates superior classification performance and efficiency compared to a simple CNN, Deep CNN, and other state-of-the-art architectures such as AlexNet, ResNet, and Inception V3 at lower computational overhead. The proposed M-CNN achieves accurate brain tumor classification with reduced overfitting risks and optimized computational efficiency. • Develop an M-CNN architecture based on the integration of multi-convolutional layers. • Analyze the impact of M-CNN's different kernel sizes on multi-class brain tumor detection. • Employ feature selection to reduce the computational complexity for brain tumor classification. • Evaluate the empirical effectiveness and computational efficiency of the proposed M-CNN with the existing CNN-based DL architectures. • Provided a comprehensive analysis using different feature sets for brain tumor detection.