Predicting Pulmonary Function from Chest X-Rays Using Deep Learning: A DenseNet Approach to Estimating FEV1/FVC Z-Scores

Somin Mindy Lee, Mohamed Abdalla, Bita Behrouzi, Soyoon Lee, Po-Chih Kuo, Han-Jay Shu, Kenneth Patrick Seastedt, Leo Anthony Celi, Benjamin Fine

Abstract: Lung disease remains a critical global health challenge, affecting millions worldwide. Pulmonary function tests (PFTs) are gold standard diagnostic tools for evaluating lung health, particularly in conditions such as chronic obstructive pulmonary disease (COPD) and asthma. However, conventional PFTs are labor-intensive, rely on specialized equipment, and can be affected by patient effort. In this study, we propose a deep learning model based on the Dense Convolutional Network (DenseNet) architecture to predict pulmonary function scores directly from chest X-ray images. By integrating transfer learning, custom loss functions, enhanced DenseNet layers, and dynamic percentile calculations, our approach predicts the ratio of forced expiratory volume in one second (FEV1) to forced vital capacity (FVC) Z-scores. Our model achieves a mean absolute error (MAE) of 0.31, indicating that predictions deviate by less than one-third of a standard deviation from the true values, and a Pearson correlation coefficient of 0.98, reflecting a near-perfect linear relationship between predicted and observed scores. For patients in the lowest 10th percentile of lung function, the model attains an accuracy of 88.41%. By leveraging routinely acquired CXRs, this method offers a scalable, accessible alternative to traditional PFTs, with the potential to improve diagnostic precision, streamline clinical workflows, and enhance patient management across diverse healthcare settings.

External IDs:dblp:conf/i2mtc/Lee0BLKSSCF25