Simulated Annealing-Based Hyperparameter Optimization of a Convolutional Neural Network for MRI Brain Tumor Classification

Sofia El Amoury, Youssef Smili, Youssef Fakhri

Brain tumor classification poses significant challenges in medical imaging, largely due to the heterogeneity and structural complexity of tumors. With Magnetic Resonance Imaging (MRI) serving as a cornerstone for diagnosis, manual interpretation by radiologists is time-consuming and prone to inter-observer variability. Recent advances in deep learning, particularly through the application of Convolutional Neural Networks (CNNs), have transformed medical image analysis by enabling automated, high-accuracy feature extraction. Despite their promise, the performance of CNNs is highly contingent upon optimal hyperparameter tuning, a process that can be both computationally demanding and pivotal for model efficacy. In this study, we employ Simulated Annealing (SA), a probabilistic metaheuristic technique, to methodically optimize the hyperparameters of a CNN architecture designed specifically for classifying brain tumors from MRI scans. Our approach employs a direct representation of hyperparameters alongside an efficient perturbation strategy, facilitating a comprehensive exploration of the parameter space. Experimental evaluations conducted on an extensive MRI dataset (N = 7023 scans classified into glioma, meningioma, no tumor and pituitary) demonstrate that our SA-optimized CNN model achieves a validation accuracy of 98.15%, thereby affirming the potential of SA in enhancing the performance of deep learning systems in medical diagnostics. These findings underscore the critical role of advanced hyperparameter optimization techniques in improving diagnostic accuracy and robustness, ultimately contributing to the development of more reliable and efficient brain tumor classification systems in clinical settings.