LRR-CED: low-resolution reconstruction-aware convolutional encoder-decoder network for direct sparse-view CT image...

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Sparse-view computed tomography (CT) reconstruction aims to reduce patient radiation dose by using fewer X-ray projections.
• This reduction in data leads to significant artifacts in images reconstructed with traditional methods like filtered-backprojection (FBP).
• Existing deep learning methods often act as post-processing steps on FBP images, rather than direct reconstruction solutions.

Purpose of the Study:

• To develop a direct deep learning-based reconstruction method for sparse-view CT.
• To leverage information from low-dimensional scout images alongside sinogram data.
• To improve the accuracy and quality of CT reconstructions with reduced radiation exposure.

Main Methods:

• A novel convolutional encoder-decoder (CED) network architecture is proposed.
• The method directly learns the projection-to-image mapping.
• FBP scout images at multiple resolutions are concatenated within the decoder to enhance reconstruction.

Main Results:

• The direct deep learning approach was evaluated using Dense Blocks and U-Net architectures.
• The proposed method demonstrated the feasibility of learning a direct sinogram-to-image mapping.
• Results were benchmarked against post-processing deep learning methods (FBP-ConvNet, DD-Net) and a total variation (TV) iterative method.

Conclusions:

• This work presents a novel approach integrating sinogram and scout image data for sparse-view CT reconstruction.
• The generalization of the method across two distinct neural network architectures is demonstrated.
• The research contributes to exploring deep learning's role in various stages of the CT image reconstruction pipeline.