A Sparse-View CT Reconstruction Method Based on Combination of DenseNet and Deconvolution.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Sparse-view computed tomography (CT) enables faster scans and lower radiation doses.
• Traditional reconstruction methods like filter backprojection (FBP) struggle with artifacts from limited data.
• Iterative reconstruction algorithms offer improvements but demand significant computational resources.

Purpose of the Study:

• To develop a novel deep learning (DL) method for sparse-view CT image reconstruction.
• To improve image quality by reducing streaking artifacts and preserving structural details.
• To create an efficient reconstruction network that balances performance and computational demands.

Main Methods:

• A two-step approach combining Filter Backprojection (FBP) with a Deep Learning (DL) neural network.
• Utilizing a DenseNet and deconvolution-based network (DD-Net) for image enhancement.
• Employing shortcut connections within DD-Net to accelerate training and increase network depth.

Main Results:

• The proposed DD-Net effectively removed streaking artifacts common in sparse-view CT.
• DD-Net demonstrated superior structure preservation compared to existing state-of-the-art methods.
• Quantitative analysis showed up to an 18% increase in structure similarity and a 42% reduction in root mean square error.

Conclusions:

• The DD-Net method shows significant potential for high-quality sparse-view CT image reconstruction.
• This DL-based approach offers a promising alternative for clinical applications requiring reduced scan times and radiation exposure.
• DD-Net achieves competitive performance, highlighting the efficacy of combining DL with traditional reconstruction techniques.