A Preprocessing Pipeline for Enhanced Epidemic Detection in Medical Imaging

Abstract

Public health response relies on timely detection of epidemic respiratory diseases such as COVID-19. While deep learning on chest X-rays offers an excellent solution, model performance is generally undermined by low data quality and class imbalance. This paper introduces a large open-access collection of 24,834 chest radiographs, manually annotated into five classes: COVID-19, tuberculosis, pneumonia, pulmonary fibrosis, and normal. We recommend an end-to-end preprocessing pipeline that includes blur correction, inpainting artifact removal, and image standardization. To address class imbalance and benefit public health priority issues, we also redistribute the classes into an imbalanced binary classification problem (Epidemic vs. Non-Epidemic). We validate the effectiveness of our approach by training deep models, demonstrating that our preprocessing pipeline enhances performance by a significant amount boosting average accuracy and F1-score by 4.7 % and 6.2 %, respectively, above models trained on raw data. The paper lays a solid foundation for the development of reliable AI devices for respiratory disease diagnosis and epidemic preparedness.