Label-Efficient CT Emphysema Segmentation via Synthesis and Test-Time Training.

Abstract

CT emphysema segmentation supports objective COPD assessment, but pixel-level annotation is costly because emphysema lesions are spatially diffuse and visually heterogeneous. Existing lesion synthesis methods rarely model emphysema-specific low-attenuation parenchymal destruction, and synthetic-trained models often degrade on real CT scans. We propose a label-efficient emphysema segmentation framework built on two complementary mechanism-level components. First, Prior-guided Emphysema Synthesis (PES) generates synthetic lesions through Gaussian-based subregion selection, within-region density modulation, and prior-guided intensity sampling. Second, Restorative Contrastive Test-Time Training (ResCon-TTT) reduces the remaining domain gap by coupling multi-dimensional dynamic feature perturbation with a restoration-contrastive auxiliary objective. On the internal dataset, PES-trained UNet achieved 70.11% DSC, and ResCon-TTT further improved DSC to 73.42%. On two external datasets, ResCon-TTT achieved 72.63% and 83.35% DSC, outperforming competing TTT/TTA methods. These results suggest that emphysema-specific synthesis and feature-level test-time adaptation can reduce annotation dependence while improving cross-center robustness. The source code is publicly available at: https://github.com/smallrookie/ResCon-TTT.git.