Optimizing neural networks for healthcare data : Ensemble learning and hyperparameter optimization

Abstract

The integration of Artificial Intelligence (AI) in medical imaging has shown substantial potential, particularly in automating disease diagnosis using deep learning architectures. Chest X-ray classification has emerged as a critical application area, enabling early detection of thoracic diseases. However, standalone Convolutional Neural Networks (CNNs) are often limited in their generalization capabilities and sensitivity to suboptimal hyperparameters. This thesis investigates the performance impact of Hyperparameter Optimization (HPO) techniques specifically random search and genetic algorithms on three pretrained CNNs: DenseNet201, VGG16, and ResNet101. After optimization, Ensemble Learning (EL) strategies, including soft voting and stacking, were employed to enhance classification accuracy and model robustness. The proposed methodology, guided by findings in the existing literature, yielded notable improvements in multiclass diagnostic performance. Nevertheless, the research faced practical constraints including difficulties in acquiring suitable data and selecting the most appropriate dataset for the task, limited GPU memory, and extended training time, which influenced model stability and optimization depth