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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Accurate emphysema segmentation on CT scans is crucial for COPD assessment but is hindered by costly pixel-level annotations due to lesion heterogeneity.
• Current synthetic data generation methods often fail to capture emphysema-specific characteristics, leading to poor performance on real scans.

Purpose of the Study:

• To develop a label-efficient framework for emphysema segmentation that reduces annotation dependence and improves cross-center robustness.
• To enhance the performance of deep learning models for emphysema segmentation using specialized synthesis and test-time adaptation techniques.

Main Methods:

• Proposed a Prior-guided Emphysema Synthesis (PES) method to generate realistic synthetic emphysema lesions.
• Introduced Restorative Contrastive Test-Time Training (ResCon-TTT) to bridge the domain gap between synthetic and real CT data.
• Employed Gaussian-based subregion selection, density modulation, and intensity sampling for PES, and feature perturbation with a contrastive objective for ResCon-TTT.

Main Results:

• A UNet model trained with PES achieved 70.11% Dice Similarity Coefficient (DSC) on an internal dataset.
• ResCon-TTT further improved the DSC to 73.42% on the internal dataset.
• On external datasets, ResCon-TTT achieved DSC scores of 72.63% and 83.35%, outperforming existing test-time adaptation methods.

Conclusions:

• Emphysema-specific data synthesis and feature-level test-time adaptation significantly reduce the reliance on manual annotations for CT segmentation.
• The proposed framework demonstrates improved robustness and generalizability across different clinical centers.
• The developed methods offer a promising approach for objective COPD assessment through efficient emphysema segmentation.