HEAL: High-Frequency Enhanced and Attention-Guided Learning Network for Sparse-View CT Reconstruction

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• X-ray computed tomography (CT) is crucial for clinical diagnosis but involves ionizing radiation risks.
• Reducing radiation dose in CT is a major research focus, with sparse-view imaging being a key strategy.
• Deep learning methods show promise for sparse-view CT reconstruction, but detail recovery under ultra-sparse conditions remains difficult.

Purpose of the Study:

• To develop an advanced deep learning network for high-quality image reconstruction in ultra-sparse-view CT.
• To address the challenge of efficiently recovering fine image details when using minimal X-ray projections.
• To improve the accuracy and detail preservation in low-dose CT imaging.

Main Methods:

• Proposed a novel deep learning network named HEAL (high-frequency enhanced and attention-guided learning Network).
• Implemented a dual-domain progressive enhancement module with fidelity and consistency constraints.
• Incorporated channel and spatial attention mechanisms for improved feature scaling.
• Introduced a high-frequency component enhancement regularization term integrating residual learning and direction-weighted total variation.

Main Results:

• The HEAL network demonstrated significant advantages in reconstruction accuracy and detail enhancement.
• Evaluated under ultra-sparse configurations (60 and 30 views), HEAL outperformed existing methods.
• The proposed optimization strategies effectively distinguished between noise and textures, preserving image details.

Conclusions:

• HEAL effectively addresses the challenge of image detail recovery in ultra-sparse-view CT reconstruction.
• The network offers a promising solution for reducing radiation dose while maintaining high image quality in CT imaging.
• This approach has the potential to enhance diagnostic capabilities in clinical settings with reduced patient exposure.