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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiovascular diseases are a leading cause of death globally.
• Invasive Coronary Angiography (ICA) is crucial for cardiac interventions.
• Automated vessel segmentation using deep learning aids stenosis assessment but faces domain shift issues due to data variations.

Purpose of the Study:

• To develop a robust Single-source Domain Generalization (SDG) method for coronary vessel segmentation.
• To overcome limitations of augmentation-based SDG methods that risk overfitting.
• To enhance the generalization capability of deep learning models across diverse clinical settings.

Main Methods:

• Proposed a progressive and targeted channel dropout method targeting the first layer of Convolutional Neural Networks (CNNs).
• Identified and progressively dropped domain-specific channels that overfit to training source features.
• Developed an architecture-agnostic method integrable with any CNN backbone.

Main Results:

• Demonstrated improved out-of-distribution performance across five diverse ICA datasets.
• Maintained in-domain performance, indicating robust generalization.
• Showcased the method's ability to stabilize feature learning and promote domain-invariant representations.

Conclusions:

• The proposed channel dropout method enhances the generalization of coronary vessel segmentation models.
• This approach provides a stable foundation for clinical deployment of AI in cardiac imaging.
• Facilitates reliable 3D reconstruction and hemodynamic analysis of coronary arteries.