Sequential vessel segmentation via deep channel attention network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease Diagnosis

Background:

• Accurate segmentation of coronary vessels in X-ray coronary angiography (XCA) is crucial for diagnosing and treating coronary artery disease.
• Challenges in XCA image segmentation include overlapping structures, low contrast, and complex background artifacts.

Purpose of the Study:

• To develop a novel deep network for accurate 2D vessel segmentation from XCA image sequences.
• To improve the robustness and performance of automated vessel segmentation in challenging XCA images.

Main Methods:

• A novel encoder-decoder deep network architecture utilizing temporal-spatial feature extraction.
• Integration of feature fusion via skip connections and a channel attention mechanism in the decoder.
• Employment of 3D convolutional layers for hierarchical feature extraction and Dice loss for training to handle class imbalance.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art algorithms in quantitative metrics and visual validation.
• Effective discrimination of vessel features from complex backgrounds was achieved.
• The network successfully segmented 2D vessel masks from the current frame using contextual frames.

Conclusions:

• The developed deep network offers a significant advancement in automated vessel segmentation for XCA images.
• The method provides a robust solution for diagnosing coronary artery disease by improving segmentation accuracy.
• Public release of the dataset and source code facilitates further research in the field.