B78-19 CT-derived Lung Deformation Metrics Improve Classification of Longitudinal Lung Function Trajectories

Rationale

Traditional CT metrics, such as emphysema quantification and airway wall thickness, capture structural lung abnormalities but provide limited information about regional lung mechanics. Jacobian determinants derived from inspiratory-expiratory CT registration quantify regional lung expansion and contraction and may capture functional impairment beyond static CT metrics. This study evaluates whether Jacobian deformation measures add value beyond traditional CT metrics for classifying spirometrically defined lung function trajectories.

Methods

Chest CT data from COPDGene were analyzed. A total of 728 individuals were selected using propensity score matching to balance age, height, sex, and race across spirometric disease categories. Lung lobes were automatically segmented using the deep learning model TotalSegmentator. Inspiratory and expiratory CT scans were non-rigidly registered using the deep learning model uniGradICON, yielding voxel-wise deformation fields. Jacobian determinants from these deformation fields were computed to quantify local lung volume change between inspiration and expiration, with lower values indicating greater lung contraction. Mean and standard deviation of the Jacobian determinant were summarized per lobe and for the whole lung to capture overall deformation and spatial variability. Lung function trajectory assignments derived from longitudinal spirometry were treated as a six-class outcome. Analyses were restricted to individuals with complete trajectory and imaging data and were stratified by sex. Multinomial logistic regression with L2 regularization modeled the six-class outcome using baseline clinical variables and traditional CT metrics, with Jacobian measures added in an extended model. Model performance was evaluated using stratified 5-fold cross-validation and macro-averaged F1 score. The added predictive value of Jacobian measures was quantified by paired differences in macro-F1 across folds, and variable importance was summarized using the L2 norm of regression coefficients.

Results

In males (N = 184), baseline and traditional CT metrics achieved a mean macro F1 of 0.32. It increased to 0.36 with Jacobian measures, corresponding to an average improvement of 0.04 ± 0.10 across folds. In females (N = 375), baseline and traditional metrics had a mean macro F1 of 0.30, improving to 0.37 with Jacobian measures, an average increase of 0.07 ± 0.05 across folds. Fold-specific differences ranged from -0.12 to 0.15 in males and 0.004 to 0.12 in females.

Conclusions

Jacobian deformation measures improved discrimination of longitudinal lung function trajectories beyond traditional CT metrics. Performance improvements were observed in both sexes, with more consistent improvements in females. These findings suggest that regional lung mechanics captured by Jacobian measures may enhance trajectory-based phenotyping beyond static structural CT features.

This abstract is funded by: None