CT image super-resolution reconstruction based on global hybrid attention

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Traditional Computer Tomography (CT) images often suffer from blurry edges and unclear textures, hindering accurate radiological interpretation.
• Existing deep learning super-resolution methods struggle to effectively map multi-level CT image details and prioritize regions of interest (ROIs).

Purpose of the Study:

• To develop an advanced CT image super-resolution network that overcomes limitations of current methods.
• To enhance the clarity and diagnostic value of CT images through improved feature extraction and fusion.

Main Methods:

• Employed stacked Swin Transformer blocks as a backbone for initial feature extraction from degraded CT images.
• Introduced a multi-branch hierarchical self-attention module (MHSM) to adaptively map and relate multi-level image features.
• Integrated a multidimensional local topological feature enhancement module (MLTEM) to refine spatial and channel features, prioritizing ROIs.

Main Results:

• The proposed network effectively extracts and fuses multi-level features from CT images.
• The MLTEM module successfully enhances features in regions of interest while suppressing background noise.
• Experimental results show superior performance compared to state-of-the-art methods in CT image super-resolution, evidenced by higher PSNR and SSIM values.

Conclusions:

• The proposed hybrid attention and global feature fusion network significantly improves CT image super-resolution.
• This advancement offers enhanced diagnostic accuracy by providing clearer and more detailed medical images.
• The method demonstrates a promising approach for medical image restoration in clinical settings.