CR4Re: Reformulating binary COPD classification using contrastive representations

Classifying heterogeneous diseases such as Chronic Obstructive Pulmonary Disease (COPD) from medical images remains a challenging task due to their diverse visual manifestations. In this paper, we introduce TopoAnomaly, a novel framework that reformulates COPD classification as an anomaly detection problem enhanced by topological data analysis (TDA). Our method first employs self-supervised contrastive learning to extract robust representations from lung patches. Then, using persistent homology, we capture multi-scale topological features from these learned embeddings. By modeling the distribution of topological signatures from healthy lung tissues via a generative model, our approach detects COPD-related anomalies as topological outliers. Experimental results show that TopoAnomaly outperforms existing state-of-the-art methods, achieving AUROC scores of 91.0% and 89.0% on two public datasets. Moreover, it produces interpretable spatial anomaly maps with inherent topological meaning, which are particularly useful for early-stage COPD detection. Evaluations under simulated real-world prevalence scenarios further demonstrate the robustness and generalizability of our method, establishing a mathematically grounded framework applicable to various medical imaging tasks.