A topology-aware visual fusion framework for patient-level emphysema classification from chest CT images

Emphysema classification from chest computed tomography (CT) remains challenging because disease-related abnormalities involve both visual appearance and structural disruption of the lung parenchyma. We propose a patient-level visual–topological deep learning framework for emphysema classification from a single 2D chest CT image per patient. A newly curated single-center dataset of 1108 anonymized chest CT images was collected in collaboration with expert radiologists. Each image corresponded to a unique patient and was labeled as normal, panlobular emphysema, paraseptal emphysema, or centroacinar emphysema. The framework integrates DINOv2-derived visual embeddings with persistent-homology-derived Betti-curve descriptors using a topology-guided cross-attention fusion module. On the held-out patient-level test set, the model achieved a macro F1-score of 0.9269 and a one-vs-rest ROC-AUC of 0.9830 for four-class classification. In the binary normal-vs-emphysema task, it achieved an ROC-AUC of 0.9942 and a PR-AUC of 0.9976. Stratified ten-fold patient-level cross-validation further confirmed the robustness of the proposed framework, yielding a mean macro F1-score of 0.9217 ± 0.0360 and a mean one-vs-rest ROC-AUC of 0.9917 ± 0.0054 for four-class emphysema classification. Ablation experiments showed that Betti-curve features were independently discriminative and that their integration with a topology encoder and cross-attention fusion improved performance over the DINOv2 image-only baseline on the same patient-level split. These results indicate that combining DINOv2 embeddings with Betti-curve descriptors yields complementary gains for patient-level emphysema subtype recognition. The anonymized dataset and source code are available at: https://github.com/muhammedtalo/Emphysema .