Related Experiment Video
Updated: Sep 28, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
538
The auto segmentation for cardiac structures using a dual-input deep learning network based on vision saliency and
Jing Wang1, Shuyu Wang1, Wei Liang2
1Department of Electric Information Engineering, Shandong Youth University Of Political Science, Jinan, China.
Journal of Applied Clinical Medical Physics
|April 1, 2022
Summary
This study introduces a deep learning model using visual attention and transformers for accurate cardiac structure segmentation in coronary CT angiography (CCTA) images, achieving high accuracy and potential for clinical use.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiovascular Imaging
Background:
- Accurate segmentation of cardiac structures in coronary CT angiography (CCTA) is essential for detailed morphological analysis, precise measurements, and functional assessments.
- Current segmentation methods may lack the precision required for comprehensive cardiac evaluation.
Purpose of the Study:
- To develop and evaluate an innovative deep learning method for automatic segmentation of cardiac structures in CCTA images.
- To assess the practical application value of the proposed segmentation technique.
Main Methods:
- A dual-input deep learning network, the Vision Attention and Transformer (VST) model, was developed incorporating a self-attention mechanism.
- The VST model was trained on 60 CCTA patient datasets with manual contours serving as ground truth, utilizing a deep supervision strategy and a combined Dice and cross-entropy loss function.
- Segmentation performance was quantitatively evaluated using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD), with statistical comparison of segmented volumes against manual segmentations.
Main Results:
- The VST model demonstrated high segmentation accuracy across various cardiac structures, with average DSC of 0.92 and average HD of 7.2 ± 2.1 mm.
- Specific DSC values included: Left Ventricular Myocardium (0.87), Left Ventricle (0.94), Left Atrium (0.90), Right Ventricle (0.92), Right Atrium (0.91), and Aorta (0.96).
- Volume comparisons showed no significant statistical difference for most structures, indicating reliable segmentation accuracy comparable to manual annotations.
Conclusions:
- The dual-input, visual saliency, and transformer-based architecture effectively segments cardiac structures with high sensitivity and specificity.
- This deep learning approach significantly improves the accuracy of automatic cardiac substructure segmentation in CCTA images.
- The method shows strong potential for enhancing the analysis and evaluation of cardiovascular morphology and function.
