Abstract
This paper develops a novel encoder-decoder deep network architecture which exploits the several contextual frames of 2D+t sequential images in a sliding window centered at current frame to segment 2D vessel masks from the current frame. The architecture is equipped with temporal-spatial feature extraction in encoder stage, feature fusion in skip connection layers and channel attention mechanism in decoder stage. In the encoder stage, a series of 3D convolutional layers are employed to hierarchically extract temporal-spatial features. Skip connection layers subsequently fuse the temporal-spatial feature maps and deliver them to the corresponding decoder stages. To efficiently discriminate vessel features from the complex and noisy backgrounds in the XCA images, the decoder stage effectively utilizes channel attention blocks to refine the intermediate feature maps from skip connection layers for subsequently decoding the refined features in 2D ways to produce the segmented vessel masks. Furthermore, Dice loss function is implemented to train the proposed deep network in order to tackle the class imbalance problem in the XCA data due to the wide distribution of complex background artifacts. Extensive experiments by comparing our method with other state-of-the-art algorithms demonstrate the proposed method's superior performance over other methods in terms of the quantitative metrics and visual validation.
Introduction
X-ray coronary angiography (XCA) remains the clinical gold standard for diagnosing coronary artery disease. However, the automated segmentation of coronary vessels from XCA sequences is significantly challenged by low contrast, high noise levels, and overlapping anatomical structures such as the spine and ribs. Furthermore, dynamic artifacts caused by cardiac and respiratory motion complicate the task for traditional 2D segmentation methods, which often fail to distinguish between moving vessels and static background noise.
Methodology
To address these challenges, we propose SVS-net, a sequential vessel segmentation network that leverages temporal-spatial context. The architecture follows an encoder-decoder structure with several key innovations designed to handle the unique properties of XCA data.
Temporal-Spatial Encoder
The encoder utilizes 3D convolutional layers (Conv3D) and 3D residual blocks (Block3D) to process a sliding window of four successive frames. This allows the network to extract hierarchical features that capture both spatial vessel structures and their temporal motion patterns, effectively treating motion as a feature rather than a source of noise.
Feature Fusion and Channel Attention
A Feature Fusion Operation (FFO) is employed in the skip connections to bridge the 3D encoder and 2D decoder. By using a 4x1x1 convolution, the temporal dimension is collapsed, distilling motion-aware features into 2D maps. In the decoder stage, Channel Attention Blocks (CAB) are integrated to refine these features. The CAB mechanism uses high-level semantic information to weight low-level feature channels, suppressing background noise and sharpening vessel structures.
Optimization
The network is trained using a Dice loss function to mitigate the severe class imbalance inherent in XCA data, where vessel pixels constitute only a small fraction of the total image area compared to the background.
Experimental Results
The proposed method was evaluated on a clinical dataset of 120 XCA sequences from Ren Ji Hospital, with 323 high-quality expert-annotated samples used for validation.
Quantitative Performance
SVS-net demonstrated superior performance compared to state-of-the-art 2D models. It achieved a Detection Rate (DR) of 0.8424 and an F-measure of 0.8428, outperforming 2D baselines by over 10% in detection rate. Ablation studies confirmed that the transition from 2D to 3D features provided a ~2.5% improvement in F-measure, while the addition of CAB contributed a further ~2% boost.
Clinical Relevance and Speed
The model achieved a significantly lower Gland Volume Error (GVEs) of 9.74% compared to baseline methods, indicating a more accurate estimation of the total vessel network volume. Furthermore, the inference speed of 0.178 seconds per frame on an NVIDIA 1080 Ti supports its potential for real-time clinical application.
Conclusion
SVS-net introduces a robust framework for sequential vessel segmentation by effectively utilizing temporal context and channel attention. By treating XCA sequences as 2D+t data, the model successfully overcomes the limitations of static 2D segmentation, providing a more accurate and noise-resistant tool for clinical diagnosis. Future work will focus on improving the segmentation of thin vessel branches through weighted loss functions.
Summary updated