Optimizing Neural Networks for Brain Tumor Detection in MR Images Using Evolutionary Computing

Evolutionary computation (EC) is a family of bio-inspired metaheuristic algorithms well-suited for global optimization in high-dimensional spaces. This paper investigates the application of EC methods to optimize convolutional neural networks (CNNs) for brain tumor detection in magnetic resonance imaging (MRI). Three EC algorithms, Particle Swarm Optimization (PSO), Moth-Flame Optimization (MFO), and its nonlinear chaotic extension with Lévy flights (NLCMFO), were applied at different stages of the pipeline, including image preprocessing, CNN hyperparameter tuning, and loss function optimization. A dataset of 3,509 MRI scans was used, comprising both T1- and T2-weighted images. The results show that the EC-based hyperparameter tuning substantially improved the baseline CNN, increasing the F1 score from 71% to 97%. While EC-assisted skull stripping produced mixed outcomes, a simple contour-based zooming method proved to be more reliable for this dataset. NLCMFO exhibited the most consistent optimization behavior, delivering stable convergence and performance comparable to a fine-tuned VGG-16 model, despite the lower architectural complexity of CNN. These findings highlight the potential of EC techniques to adapt lightweight custom CNNs to domain-specific tasks, demonstrating that EC-optimized simple architectures can achieve performance rivaling state-of-the-art pre-trained models.