Abstract Rationale Lung shape is an emerging biomarker that has been shown to vary with age and disease. Prior studies using statistical shape modelling have shown that whole-lung geometry relates to chronic obstructive pulmonary disease (COPD) severity. However, the relationship between lobe-specific lung shape and lung function, and how it reflects tissue microstructure, remains unknown. Because macroscopic lung and lobe morphology arise from - and signal - microstructural remodelling, we hypothesize that lobe-level shape in COPD is associated with loss of lung function, and that this relationship would be strengthened when accounting for local tissue heterogeneity. Methods Shape-aware conditional generative modelling has previously been used to analyse other anatomical structures. This method uses low-dimensional shape representation and a conditioning vector that contains other field-based information to train a conditional generative model. We developed LobeSDFNet, a shape-aware conditional generative model that learns a nonlinear representation of lung and lobe shape from volumetric signed distance function (SDFs). The model was conditioned on quantitative measures of tissue heterogeneity derived from normal-appearing parenchyma on inspiratory CT, representing local microstructure. The latent vectors from the model captured low-dimensional nonlinear shape variation, while reconstruction accuracy was evaluated with and without the tissue heterogeneity condition. Results We tested the approach using end-inspiration lobe shape and tissue heterogeneity in 4,160 participants distributed across GOLD stages 1-4. Our analysis has revealed significant associations between shape (represented by latent vectors) and percent predicted FEV1. SHAP analysis identified latent features 12, 29, 20 and 31 as strongest contributures to model output (mean|SHAP| 1.0), indicating these shape embedded components most strongly influenced FEV1%predicted. Among them, latent 7 showed a small but statistically significant positive mean SHAP (p = 0.02), suggesting a consistent relationsip between this latent dimension and better preserved lung function. Although most influential latents were not directionally significant, their large SHAP magnitudes indicated consistent importance for prediction across subjects, capturing structural variance critical to model performance. We have also confirmed that shape reconstruction is significantly more accurate when tissue heterogeneity is included in the model. Conclusions The conditional generative model learns a compact, biologically informed representation of lung shape that encodes dependence on tissue microstructure. A multi-lobar analysis helps to better understand lobe-level shape and heterogeneity changes with stage of disease. The latent space can potentially be used in biomarker analysis to better stage the disease severity of COPD. This abstract is funded by: This work was supported by NHLBI grants U01 HL089897 and U01 HL089856 and by NIH contract 75N92023D00011.
John et al. (Fri,) studied this question.