What question did this study set out to answer?

The aim is to evaluate if deep learning improves quantification of lung density and opacities on chest radiographs compared to traditional methods.

May 20, 2026

A24-09 Deep Learning for Automated Lung-Specific Density Quantification in Chest Radiographs of Acute Respiratory Failure

Key Points

The aim is to evaluate if deep learning improves quantification of lung density and opacities on chest radiographs compared to traditional methods.
Analyzed 5,690 chest x-rays from acute respiratory failure patients.
Trained two deep learning models: a regression-based Siamese network and a classification-based DenseNet.
Model performance assessed via correlation and reliability metrics against physician annotations.
DenseNet showed Spearman correlations of 0.84 for both lungs with F1-scores of 0.73 and 0.70.
Siamese network demonstrated superior performance with correlations of 0.86 (right) and 0.84 (left), ICC of 0.91 and 0.90 respectively.
Both models effectively distinguished between severity categories with clinically relevant accuracy.

Abstract

Abstract Rationale Quantifying the extent and density of radiographic opacities on chest x-rays (CXRs) is critical for diagnosis and management of acute respiratory distress syndrome (ARDS). The Radiographic Assessment of Lung Edema (RALE) score provides semi-quantitative measurement of opacity extent and density, correlating with clinical outcomes including mortality and ventilator-free days. However, aggregating bilateral lung findings into a single composite RALE score may obscure lateralized disease patterns and limit prognostic precision. Additionally, subjective density estimation introduces inter-rater variability. We hypothesized that lung-specific deep learning quantification would provide more granular and reproducible radiographic assessment than traditional composite RALE scoring. Methods We analyzed 5,690 CXRs from patients with acute respiratory failure, representing one of the largest and more rigorously annotated RALE datasets available. Physicians underwent iterative training with feedback to ensure reliable quadrant-specific annotation of density (0-3) and extent (0-4) scores. Each lung was classified into four severity categories (healthy, low, medium, high) based on maximum quadrant density and extent values. We trained two Convolutional Neural Network (CNN) architectures: a regression-based Siamese network Li, MD, et al, 2020 and a classification-based DenseNet Huang, G, et al, 2017. Model performance was evaluated against physician annotations using Spearman correlation, intraclass correlation coefficients (ICC), and F1-scores. Analysis was performed separately for right and left lungs to capture lateralized pathology. Results The DenseNet classifier demonstrated strong Spearman correlation for both lungs (R = 0.84 right and left) with F1-scores of 0.73 (right) and 0.70 (left) across severity categories. The Siamese regression network had superior performance with Spearman correlation of 0.86 (right) and 0.84 (left) and excellent inter-rater reliability (ICC2,k = 0.91 right, ICC2,k = 0.90 left), approaching human expert agreement levels. Both models successfully discriminate between severity categories with clinically meaningful accuracy. Conclusion Deep learning models can reliably quantify lung-specific radiographic opacity with performance approaching inter-physician agreement. Lung-lateralized assessment preserves spatial information lost in composite RALE scores, potentially improving phenotyping of asymmetric lung disease and longitudinal monitoring. Automated, reproducible quantification may enhance clinical workflow efficiency, reduce inter-observer variability, and provide standardized metrics for treatment response assessment. Future validation should examine whether lung-specific density quantification improves prediction of clinical outcomes including ventilator days, ICU length of stay, and mortality in ARDS populations. This abstract is funded by: None

Mark Helpful

Bookmark

Relay