Domain-Specific Progressive Channel Dropout: Single-Source Domain Generalization for Vessel Segmentation in X-ray Coronary Angiography.

Cardiovascular diseases remain the leading global cause of death, with Invasive Coronary Angiography (ICA) being the gold standard for cardiac interventions. While deep learning enables automated vessel segmentation from ICA for critical tasks like stenosis assessment, these models struggle with domain shifts across clinical settings due to variations in protocols, equipment, and patient demographics. This is exacerbated by the scarcity of annotated datasets, making Single-source Domain Generalization (SDG) essential. Current SDG methods are augmentation-based and risk overfitting to synthetic variations. To address this, we propose a progressive and targeted channel dropout method that explicitly targets channel behavior in the first layer of a Convolutional Neural Network (CNN). By identifying and progressively dropping domain-specific channels that overfit to training source features, our proposed method stabilizes feature learning and promotes domain-invariant representations. Our architecture-agnostic method can be integrated with any CNN backbone to enhance generalization. Extensive evaluation across five diverse ICA datasets demonstrates improved out-of-distribution performance while maintaining in-domain performance, establishing robust foundation for clinical deployment.Clinical relevance- Our proposed method aids reliable deployment of coronary vessel segmentation models for x-ray angiography by improving generalization across diverse imaging conditions, thus supporting important downstream tasks, such as 3D reconstruction of coronary arteries and non-invasive hemodynamic analysis.