Evaluating Lightweight Vision Transformers for Chest Disease Detection in Low-Resource Clinical Settings

This study presents a comprehensive evaluation of lightweight Vision Transformers (ViTs) for chest disease detection in low-resource clinical settings using the Indiana Chest X-ray Reports dataset. The research addresses the critical need for computationally efficient diagnostic models that can operate effectively in resource-constrained healthcare environments. Through systematic preprocessing and exploratory data analysis, 349 validated normal chest X-ray cases were analysed across ten distinct radiological finding categories. The dataset demonstrated excellent characteristics for training lightweight ViTs, with a balanced distribution (coefficient of variation = 0.47) and strong clinical indication–finding correlations (r > 0.6, p < 0.05). Performance evaluation revealed that lightweight Vision Transformers achieved promising diagnostic accuracy, with precision scores reaching 1.0, while maintaining computational efficiency suitable for deployment in low-resource settings. The keyword extraction algorithm successfully identified medical conditions with 100% data quality assurance following preprocessing. Statistical analysis confirmed dataset suitability for machine learning applications, with comprehensive terminological diversity (847 unique medical terms) and clinically meaningful correlations between indications and findings. The findings demonstrate that lightweight Vision Transformers represent a viable solution for addressing diagnostic challenges in resource-limited healthcare environments, offering an optimal balance between diagnostic accuracy and computational efficiency. This research contributes to the broader goal of democratising AI-powered healthcare solutions and reducing diagnostic disparities in underserved populations worldwide.