AngioDG: Interpretable Channel-informed Feature-modulated Single-source Domain Generalization for Coronary Vessel Segmentation in X-ray Angiography

TL;DR

AngioDG addresses domain shifts in coronary vessel segmentation from X-ray angiography by targeting the contributions of early feature channels. It combines instance-wise channel whitening in the first convolutional layer with a channel-importance estimation and a Weighted Channel Attention (WCA) module during fine-tuning to emphasize domain-invariant channels while suppressing domain-specific ones. Evaluated across six XCA datasets, AngioDG achieves the best out-of-distribution performance on multiple targets while maintaining in-domain accuracy, and provides interpretable insights into which channels drive generalization. This approach offers a principled, interpretable SDG solution for reliable vessel segmentation that supports downstream tasks like stenosis assessment and 3D vascular reconstruction in real-time interventions.

Abstract

Cardiovascular diseases are the leading cause of death globally, with X-ray Coronary Angiography (XCA) as the gold standard during real-time cardiac interventions. Segmentation of coronary vessels from XCA can facilitate downstream quantitative assessments, such as measurement of the stenosis severity and enhancing clinical decision-making. However, developing generalizable vessel segmentation models for XCA is challenging due to variations in imaging protocols and patient demographics that cause domain shifts. These limitations are exacerbated by the lack of annotated datasets, making Single-source Domain Generalization (SDG) a necessary solution for achieving generalization. Existing SDG methods are largely augmentation-based, which may not guarantee the mitigation of overfitting to augmented or synthetic domains. We propose a novel approach, ``AngioDG", to bridge this gap by channel regularization strategy to promote generalization. Our method identifies the contributions of early feature channels to task-specific metrics for DG, facilitating interpretability, and then reweights channels to calibrate and amplify domain-invariant features while attenuating domain-specific ones. We evaluate AngioDG on 6 x-ray angiography datasets for coronary vessels segmentation, achieving the best out-of-distribution performance among the compared methods, while maintaining consistent in-domain test performance.

Figures (3)

• Figure 1: Proposed AngioDG for SDG in coronary vessels segmentation from x-ray angiograms. (a) We use a U-Net with a ResNet-50 encoder and a 1024-channel decoder bottleneck (U-Net-R50-1k) as the model architecture. Feature maps (pre-BN) from the first convolutional layer (denoted as $\textbf{X}$ in (a)) are used in (b)-(d). (b) For initial training, in addition to the main segmentation loss, an annealed channel whitening loss is used to promote decorrelation of feature channels in the first layer. (c) Using pre-trained model weights from (b), we obtain channel importance by dropping channels and measuring the impact on Dice and clDice. Channels are weighted based on how their removal changes the Dice and clDice compared to the baseline scores obtained on the validation set. (d) In the final fine-tuning phase, we include WCA module after the first layer (pre-BN) using the channel-importance matrix $\mathbf{D}^{\mathbf{w}}$ (initialized from (c) and fine-tuned) to modulate contributions, in order to improve DG by emphasizing domain-invariant channels while suppressing domain-specific ones.
• Figure 2: Visualization of channel importance. We visualize each channel by overlaying the top $5\%$ of each channel’s activations ($\geq 95$th percentile) in white in (b) and (c): (a) relative percentage change in Dice (Eq. \ref{['eq:delta_dice']}) and clDice (Eq. \ref{['eq:delta_cldice']}) scores on the internal validation set after dropping each channel; (b) domain-invariant channels that capture vessel structures, shown in green dashed lines in (a); (c) domain-specific channels that often capture noise or shadows, shown in orange dashed lines in (a).
• Figure 3: Qualitative performance analysis of coronary vessels segmentation across OOD datasets for ablation components, AngioDG, MixStyle, and EFDMix. True positive, false positive, and false negative are shown in white, blue, and red, respectively. The incorporation of different components in AngioDG demonstrates improvements in vessel delineation and generalization.