Emphysema can progress unevenly across the lungs, with similar overall density decline masking different regional patterns. A 2026 publication in Radiology evaluates a CT-based deep learning model that predicts lobe-specific emphysema progression in people with and without COPD. Using baseline chest CT scans from COPDGene and ECLIPSE, the model forecasts annualised changes in volume-adjusted lung density and compares predicted regional decline with observed changes over time.
A Regional View of Emphysema Change
Emphysema creates a heterogeneous pattern of damage across the lungs, which can affect assessment and treatment planning. Regional disease burden is especially relevant when care depends on knowing where lung destruction is most pronounced. Existing CT-based density measurements provide objective and reproducible information about emphysematous tissue, but they generally support global assessment rather than prediction of how each lobe may change.
The model takes a lobe-based approach. It analyses baseline chest CT scans and predicts future density decline in the right upper, right middle, right lower, left upper and left lower lobes. The outcome centres on annualised change in volume-adjusted lung density, a measure used to assess emphysema progression.
CT images came from large longitudinal cohorts that included participants with COPD and participants with normal lung function. COPDGene supplied the training and internal testing data, while ECLIPSE supplied external testing. Imaging data underwent harmonisation to reduce differences caused by scanner protocols, reconstruction methods, dose and image noise. The framework learns local lung structure from small CT image regions, summarises those features into lobe-level profiles and uses a transformer-based model to predict density decline for each lobe.
Must Read: CT-Based Model Aids Lung Lesion Diagnosis
Predictions Align with Observed Decline
Model predictions showed a positive relationship with observed lobar density decline in both the internal and external test sets. Performance remained consistent across the two cohorts, despite differences in participant characteristics and CT acquisition. Agreement testing showed that predicted and observed values generally aligned, although external testing showed greater variability than internal testing.
Performance varied by lung region. Upper lobes produced stronger results than lower lobes across the datasets. The left upper lobe showed the strongest performance in the COPDGene internal test set, while the right upper lobe performed best in ECLIPSE. Lower performance appeared in the right middle lobe in COPDGene and the left lower lobe in ECLIPSE.
Weighted aggregation of lobar predictions also produced whole-lung estimates of annualised lung density decline. These whole-lung estimates correlated with observed values and improved overall performance compared with lobe-level averages. The deep learning model outperformed a baseline clinical model built from demographic characteristics, smoking history and CT density measurements. The regional approach therefore provides lobe-specific information while also supporting global assessment when lobar predictions are combined.
Factors That Shape Prediction Accuracy
The model also classified lobes at risk of accelerated emphysema progression using predicted annualised density decline and prespecified thresholds. Internal and external testing produced similar classification performance, with upper lobes again showing stronger results. The left upper lobe consistently performed well across risk thresholds, while right middle and lower lobes performed less strongly but remained within a moderate predictive range. Specificity generally exceeded sensitivity, especially in upper lobes.
Performance differed across participant subgroups. Participants with baseline emphysema showed stronger classification results than those without baseline emphysema. Prediction of new progression among participants without baseline emphysema was weaker but still measurable across lobes. Participants who currently smoked showed slightly higher results than those who formerly smoked, particularly in the upper lobes.
Body mass index, total lung capacity, scanner characteristics and disease distribution also influenced accuracy. Lower body mass index was associated with stronger classification performance, while higher body mass index coincided with greater CT image noise and lower total lung capacity. Higher image noise increased model error, and higher total lung capacity was associated with better accuracy. Scanner-related variability remained visible after harmonisation, particularly in lower lobes. The model also performed less strongly in participants with uniform emphysema distribution than in those with upper-lobe or lower-lobe predominance.
The CT-based framework predicts lobe-specific lung density decline and identifies lobes with accelerated emphysema progression in participants with and without COPD. Predictions align with observed regional decline in internal and external testing, with stronger performance in upper lobes and in participants with established emphysema. The deep learning approach performs better than a baseline clinical model built from demographic, smoking and CT density measures. Accuracy varies with body mass index, total lung capacity, scanner characteristics and disease distribution, making image quality and patient characteristics important considerations for regional emphysema prediction.
Source: Radiology
Image Credit: iStock
References:
Curiale AH, Niethammer M & Estépar RSJ (2026) Prediction of Lobar Emphysema Progression with a CT-Based Foundational Model. Radiology; 319:2.
