General CNN model for biomedical image classification via genetic algorithm-based hyperparameter optimization

This study addresses the challenge of hyperparameter selection, a key factor affecting convolutional neural networks (CNNs) performance in biomedical image classification. A genetic algorithm (GA) is employed to optimize activation function, padding, number of filters, kernel size, dropout rate, pooling size, and batch size. The optimized CNN is trained on brain Magnetic Resonance Imaging (MRI) images of Multiple Sclerosis (MS) and validated on Alzheimer’s MRI and COVID-19 chest X-ray datasets. Results show substantial improvements across all datasets. On MS, the proposed model achieves up to 37.6 % F1-score and 33.7 % accuracy gains compared to other models. On Alzheimer’s, improvements reach 32.8 % in F1-score and 32.5 % in accuracy. For COVID-19, gains are smaller but consistent, ranging from 0.8 % to 12.1 %. Overall, the GA-optimized CNN consistently outperforms widely used architectures such as Xception, InceptionV3, VGG16, VGG19, AlexNet, ResNet50, and GoogleNet, demonstrating both enhanced classification performance and strong generalizability across biomedical imaging tasks.