Designing Optimal CNNs Architectures Using Metaheuristic Algorithms Applied to the Classification of Alzheimer's Disease

Abstract

Convolutional Neural Networks are extensively utilized across various industries, proving to be highly effective for tasks such as image or video processing, pattern recognition and classification. However, the design of CNN architectures presents significant challenges, particularly in determining the optimal CNN parameters. CNN architectures comprise numerous parameters, and their configurations can produce diverse classification results when applied to the same tasks. Typically, setting hyper-parameter values involves a complex search process, often relying on random search, extensive testing, or manual adjustment. To address this challenge, this study proposes the analysis and implementation of two meta-heuristic approaches: Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) algorithms. These approaches aim to automatically design optimal CNN architectures and enhance their performance. The optimized architectures are specifically employed in the classification of neurodegenerative diseases, with a focus on Alzheimer's image datasets.