Comparative evaluation of deep neural networks for brain tumor classification from magnetic resonance imaging

Abstract

The growing prevalence of brain tumors and the complexity of magnetic resonance image interpretation create a need for automated decision support tools that can improve diagnostic accuracy and reduce the workload on medical specialists. Modern deep learning approaches demonstrate significant potential in medical image analysis, but the selection of appropriate model architectures requires systematic comparative evaluation considering accuracy, robustness, and computational efficiency. The aim of this study is to evaluate the performance of modern deep learning architectures for multiclass brain tumor classification based on magnetic resonance images and to determine their suitability for practical implementation under different computational conditions. The research is based on the application of transfer learning using several convolutional neural network architectures. The dataset was prepared through cleaning, normalization, and augmentation procedures to improve model generalization and robustness. The models were trained and evaluated using multiple quality criteria, including classification performance, stability under data distortions, and interpretability analysis based on visualization of decision regions. Computational complexity and inference efficiency were also analyzed to assess deployment feasibility. The comparative evaluation demonstrated that deep neural networks are capable of reliably distinguishing between different tumor types and normal brain conditions. One architecture showed the highest overall classification performance and robustness to noisy input data, while architecture provided a more balanced trade off between computational efficiency and prediction quality, making it suitable for resource constrained environments. Visualization analysis confirmed that the models focus on diagnostically relevant regions of the images, supporting the validity of their predictions. The study confirms the effectiveness of deep learning models for automated brain tumor classification and highlights their potential for integration into intelligent medical decision support systems. The obtained results demonstrate practical value for clinical applications, educational purposes, and further research in medical image analysis, particularly in scenarios requiring accurate and fast preliminary diagnosis.