Hybrid U-Net and Swin-transformer network for limited-angle cardiac computed tomography

Area of Science:

• Medical Imaging
• Cardiovascular Disease Diagnosis
• Deep Learning in Medical Imaging

Background:

• Cardiac computed tomography (CT) is crucial for diagnosing cardiovascular disease.
• Image quality in cardiac CT is limited by temporal resolution, noise, and artifacts, especially when using limited-angle projections to image the beating heart.
• Reconstructing high-quality cardiac CT images from limited-angle projections remains a significant challenge.

Purpose of the Study:

• To develop a method for reconstructing high-quality cardiac CT images from limited-angle projections.
• To address the challenges of image noise and artifacts inherent in limited-angle CT data.
• To improve the diagnostic performance of cardiac CT by enhancing image quality.

Main Methods:

• Proposed a hybrid deep learning model, termed U-Swin, integrating U-Net and Swin-transformer architectures.
• U-Net component is designed to restore structural information lost due to missing projection data and artifacts.
• Swin-transformer component is utilized to capture detailed global feature distributions.

Main Results:

• The U-Swin model demonstrated superior performance compared to state-of-the-art deep learning methods.
• Validation was performed using both synthetic (XCAT) and clinical (COCA) cardiac datasets.
• The proposed method effectively reconstructs high-quality cardiac CT images from limited-angle projections.

Conclusions:

• The U-Swin hybrid model offers a promising solution for high-quality cardiac CT image reconstruction from limited-angle data.
• This advancement has the potential to significantly improve temporal resolution for imaging the beating heart.
• The developed method can enhance the diagnostic capabilities of cardiac CT for cardiovascular diseases.