CNN Hyper-parameter Optimization Using Simulated Annealing for MRI Brain Tumor Image Classification

Abstract

Accurate brain tumor classification remains a pivotal challenge in medical imaging, where heterogeneous tumor morphology and subtle anatomical variations complicate diagnostic precision. While Magnetic Resonance Imaging (MRI) serves as the clinical gold standard, manual interpretation of its high-dimensional data can be challenging. Although Convolutional Neural Networks (CNNs) have demonstrated transformative potential in automating tumor analysis, their efficacy heavily depends on hyper-parameter optimization (HPO). This study introduces a novel integration of Simulated Annealing (SA), a metaheuristic optimization algorithm inspired by thermodynamic principles, to systematically optimize the hyper-parameters of a CNN for brain tumor classification using MRI images. Our framework incorporates a direct representation of hyperparameters, an efficient perturbation strategy, and a dual evaluation criteria balancing accuracy and model complexity. Experimental results showed that SA applied to CNN hyper-parameter optimization achieved a validation accuracy of 97.25%.