Abstract Rationale This work aims to enhance multi-disease lung classification and phenotyping from CT scans through the development of a contrastive learning framework that generates disease-discriminative image embeddings. Methods We recruited 1,187 participants across four groups: asthma (n = 315), COPD (n = 355), post-COVID-19 (n = 375), and healthy controls (n = 142). Imaging analyses included 1,003 single scan cases and a longitudinal subgroup of 92 participants (33 asthma, 59 post-COVID-19) with a follow-up scan. We trained a novel contrastive model with an expert-conditioned routing network to encourage disease-aware representation paths and adaptive temperature scaling to stabilize the contrastive learning objective. Classification was assessed by macro area under the ROC curve (macro-AUC). Embedding quality was examined via unsupervised k-means clustering and visualization (t-SNE). Quantitative CT (qCT) metrics differences across clusters were tested with Welch’s ANOVA. We also evaluated whether embeddings could track patient-level progression/regression in the longitudinal subgroup and predict qCT metrics via regression (R²). Results The model achieved a 4-class macro-AUC score of 0.893. Embeddings showed clear between-class separation on t-SNE and yielded clusters with significantly different qCT profiles by Welch’s ANOVA. In the longitudinal subgroup, patient trajectories in embedding space reflected clinical regression/improvement. Embedding-based prediction of qCT metrics was strongest for the Jacobian metric (R²=0.61). Conclusion An expert-conditioned contrastive learning approach produced robust, disease-discriminative CT embeddings that support accurate classification, unsupervised phenotyping, longitudinal tracking, and qCT metrics prediction. This representation may facilitate trial stratification and monitoring of disease course across asthma, COPD, and post-COVID-19 populations.FIGURE. Contrastive learning pipeline for lung disease classification. T and T’ denotes the two types of transformation and X and X’ are the two view of transformed images. The learned embeddings support supervised prediction and downstream analyses, including k-means clustering and feature-space evaluation and are later used for longitudinal trajectory tracking and qCT metrics prediction. This abstract is funded by: NIH Grant R01-HL168116, P30 ES005605, and ED P116S21000
Munira et al. (Fri,) studied this question.