A novel lightweight CNN design for MRI brain tumor image classification with performance-driven optimization

Accurately classifying brain tumors is a crucial task in medical imaging analysis, with significant implications for diagnosis and treatment planning. Early detection of brain tumors helps to save lives and ensures individuals remain in a safe zone. This research paper presents a novel lightweight custom-built Convolutional Neural Network (CNN) architecture optimized for brain tumor classification and is compared with standard pre-trained CNN architectures like VGG, Xception, and ResNet. It explores the impact of different Convolutional kernel sizes, max pooling strategies, batch sizes, and cross-dataset validation on model performance. A series of CNN models built with valid kernel size, pooling size, and batch sizes were trained and evaluated with performance metrics, including accuracy, precision, recall, and F1-score, which were analyzed across these different configurations. Additionally, five cross-dataset validations were conducted to assess model generalization. The results indicate that increasing kernel sizes and batch sizes does not improve model performance. The optimal configuration of a 4 × 4 Convolutional kernel, 4 × 4 max pooling size, and a batch size of 64 achieved the highest performance, with an accuracy of 99.54% and strong precision and recall metrics across all datasets. The proposed lightweight model, with a compact size of just 6.89 MB and only 1.8 million parameters, features a streamlined architecture of just nine layers while maintaining high accuracy. New findings suggest that efficiency in lightweight model design can be achieved with high classification accuracy, providing valuable insights for medical image analysis.