From Cross-Sectional CT to Dynamic Insights: Pseudotime-Based Modeling of Lung Nodule Progression

Abstract

Early detection of lung cancer relies on a comprehensive understanding of the progression of pulmonary nodules. Existing longitudinal modeling approaches are constrained due to the limited availability of longitudinal datasets and the failure to capture the inter-nodular relationship. In this study, we present one of the first applications of pseudotime inference, adapted from single-cell RNA sequencing studies, to reconstruct progression trajectories of nodules from cross-sectional CT images. We collected 13,626 nodule snapshots from two screening cohorts and reserved a longitudinal test set for evaluation. We compared a graph-based pseudotime method, diffusion pseudotime, and an unsupervised deep learning framework combining a variational autoencoder and a neural ordinary differential equation. Both approaches demonstrate longitudinal consistency, with malignant nodules showing a higher correlation between pseudotime and actual time. Pseudotime aligns with clinically relevant features such as irregular margins and solid consistency. Furthermore, pseudotime and delta-pseudotime effectively stratify nodules into distinct malignancy risk groups and remain significant independent predictors of malignancy after adjusting for established semantic biomarkers. Our study highlights pseudotime inference as a promising tool for dynamic modeling of lesion progression using static imaging data.