Abstract Rationale Low-dose computed tomography (LDCT) screening reduces lung cancer mortality but remains underutilized and inconsistent. Only approximately 16% of eligible high-risk adults undergo screening, and up to 24% of initial LDCT findings are false positives, driving unnecessary procedures and patient anxiety. Previous studies have shown that undesired variations in imaging, i.e., differences in reconstruction kernels and scanner field-of-view (FOV), negatively influence quantitative biomarker analyses and downstream lung cancer risk modeling. These inconsistencies limit the reproducibility of LDCT-based screening across institutions. To address this, we propose a lung CT foundation model that standardizes LDCT by harmonizing kernel and FOV variability, enabling more consistent characterization of thoracic anatomy across large-scale real-world datasets. Objective To design an AI foundation model that integrates kernel harmonization and FOV restoration within a single latent representation framework, producing standardized LDCT representations for large-scale lung imaging research and AI development. Methods We developed a latent diffusion-based foundation model using LDCT images of over 40,163 subjects from multiple (NLST, LUNA16, COVID-19Lung, LymphNode, and RIDER-Lung-CT) cohorts. The model was designed to learn an internal representation of LDCT images that captures anatomical information while reducing undesired variability caused by image acquisition (Figure1). Within this framework, two key components were incorporated: FOV restoration, which predicts missing body regions (e.g., shoulders) when the imaging does not cover the full anatomy, and kernel harmonization, which adjusts for differences in image sharpness and contrast resulting from scanner-specific reconstruction settings. The FOV restoration process explicitly distinguishes between acquired and missing anatomical information, ensuring that the completed images remain faithful to the original data. Results A tissue truncation index (TTI) was introduced to quantify truncation severity, ranging from 0% (no truncation) to 100% (fully truncated). Analysis showed that 67.2% of the training dataset had TTI 10%, indicating widespread truncation in LDCTs. Applying the proposed model effectively detruncated these scans. Quantitative evaluation showed a mean structural similarity index (SSIM) of 83.19 ± 5.61% between detruncated and reference images, while downstream body composition analysis showed improved correlations between imaging-derived fat and muscle indexes following FOV restoration. Ongoing work extends kernel harmonization to original resolution for validation. Conclusion This work presents a step toward a lung CT foundation model that mitigates acquisition variability in LDCT. By restoring full-anatomy context and moving toward kernel standardization, the framework establishes a scalable foundation for population-level harmonization and multi-organ analysis, advancing data reliability for lung cancer screening and future AI development. This abstract is funded by: This work was supported in part by the National Artificial Intelligence Research Resource (NAIRR) Pilot Project (NAIRR240468, Zuo), the National Institutes of Health (NIH) grants R01CA253923 (Landman & Maldonado), R01CA275015 (Maldonado & Lenburg), and U01CA196405 (Maldonado).
Zuo et al. (Fri,) studied this question.