B111-18 Detection and Radiomics Characterization of Pulmonary Nodules in COPDgene Dataset

Abstract Rationale Individuals with COPD have an increased risk of lung cancer, yet the radiomic characteristics of pulmonary nodules in this high-risk population remain poorly defined. Emphysema and smoking-related airway injury influence nodule biology, morphology, and malignancy potential, but large-scale imaging studies integrating automated nodule detection with quantitative phenotyping are limited. Leveraging COPDGene baseline CT scans, we sought to automatically identify and characterize pulmonary nodules and compare radiomic features across COPD status to better understand nodule phenotypes in smokers with and without airflow obstruction. Method Leveraging open-source datasets, such as Covid-19 CT Lung, Lung Image Database Consortium (LIDC), Medical Segmentation Decathlon Lung, and NSCLC-Radiomics, a residual encoder UNet was trained for 1000 epochs to detect and segment nodules. Whole chest CTs were patched into (112, 256, 256) regions with average slice thickness of 1mm. The model was applied to the COPDGene Phase 1 dataset (9419 individuals). Using pyradiomics, 44, 535 AI detected nodules were characterized. Nodule characteristics were averaged per individuals and stratified into COPD (GOLD status 0) and non-COPD (GOLD 0, -1) group. Univariate analysis was performed non-parametric statistics. Results Among 9419 COPDGene Phase 1 participants, 5244 (55. 67%) were non-COPD, while 4175 (44. 32%) showed signs of COPD (Table 1). Participants with COPD were older and had greater smoking exposure compared with non-COPD (median age 63. 2 vs. 55. 1 years; pack-years 45. 3 vs. 35. 0; both p 0. 001). AI detected nodules were identified in 78% of non-COPD and 85% of COPD participants. COPD individuals had a higher lesion burden per CTs (median 3. 0 2. 0, 6. 0) when compared to non-COPD (median 3. 0 1. 0, 4. 0; p 0. 001). AI detected nodules of COPD group were larger (median 8. 1 5. 9, 12. 4 vs. 7. 4 5. 5, 11. 0; p 0. 001) and less round (median 0. 8 0. 7, 0. 8 vs. 0. 8 0. 8, 0. 8; p 0. 001). Conclusion In COPDGene, AI detected nodules demonstrated distinct morphologic and radiomic features across disease status. Individuals with COPD had a higher nodule burden, larger and less spherical, alongside greater emphysema and gas trapping. These findings suggest that COPD-related lung injury may shape lesion phenotype, supporting the biologic link between chronic airway/parenchymal damage and lesion development. Automated lesion phenotyping at population scale provides an opportunity to refine lung cancer risk stratification in smokers and improve imaging biomarkers for malignancy surveillance in COPD. Funding K00 CA27471 and 1R01HL149877 This abstract is funded by: NCI K00CA27471; NHLBI 1R01HL149877; NHLBI U01 HL089897; NHLBI U01 HL089856; NIH Contract 75N92023D00011