Improving Chest X-ray Image Classification Efficiency through a Grid-Search Optimized Weighted Soft Voting Ensemble

Abstract

Independent deep learning models are often limited by their inherent inductive bias: convolutional neural networks (CNNs) excel at automatically extracting local textures and features but lack global contextual awareness. Vision Transformers (ViTs) can capture long-range dependencies but are highly complex and require large datasets. This paper proposes a hybrid framework that integrates the different strengths of CNNs (ResNet-50, DenseNet-201) and Transformer (Swin-T) models. Instead of using the standard soft voting method, where all models are treated equally, we introduce a weighted soft voting strategy based on a grid search. The optimal contribution of each baseline model is refined via a comprehensive grid search on a validation dataset, and the final evaluation is performed on an independent test dataset to avoid data leakage. Experiments were conducted on two chest X-ray datasets, each comprising four pathological classes. Our proposed soft voting-based improvement achieved better classification performance compared to individual baseline models and traditional ensemble methods. On the ChestXray-ImageDataSet dataset, accuracy increased by 0.23% and 3.22%, respectively, compared with the best-performing model, Soft Voting, and the worst-performing model, ResNet-50. On the COVID-19 dataset, the proposed method increased accuracy by 0.54% compared to the best-performing ResNet-50 model and 4.76% compared to the worst-performing baseline model Swin-T, using the same training method.