2025 28th International Conference on Computer and Information Technology (ICCIT)
DOI: 10.1109/ICCIT68739.2025.11491761
Abstract
Multiclass classification of brain tumors based on magnetic resonance imaging (MRI) is a challenging task due to the subtlety and variability associated with manual diagnostic protocols. The present work proposes OptiNeuroNet, a new convolutional neural network with automated hyperparameter optimization to overcome such difficulties in multiclass brain tumor classification. Unlike conventional approaches which solely depend on pre-trained architectures or manual tuning, our framework uniquely combines Optuna driven optimization with custom attention enhanced models to achieve domain-specific adaptability. Through the Optuna toolkit, kernel sizes, numbers of filters, activation functions, and dropout probabilities are systematically optimized as hyperparameters to enhance learning stability and generalizability. Comparisons are made with outstanding pre-trained models, including ResNet-50, EfficientNet-B1, MobileNetV3-Large, and Vision Transformer, using a benchmark dataset of 12,064 T1-weighted contrast-enhanced MRI images, categorized into four classes: glioma, meningioma, pituitary, and no tumor. Despite achieving considerable accuracy with EfficientNet-B1, the NeuroCBAM model performed best among custom models in terms of generalization performance, utilizing attention mechanisms, and yielded 95\% accuracy with steadily stable validation trends, low overfitting, and efficient convergence. The results show that NeuroCBAM not only closely competes with outstanding pre-trained networks, but it also provides a tuned, domain-specific solution balancing accuracy, efficiency, and robustness. The integration of automated optimization with systematic benchmarking overcomes the associated limitations of generic pre-trained architectures, opening a reliable path towards computer-aided diagnostic systems that are applicable in a clinical scenario, such as in brain tumor classification.