Sinogram Interpolation Inspired by Single-Image Super Resolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) uses radiation scans to create images of internal structures.
• High radiation doses and limited views can compromise image quality and patient safety.
• The sparse-view problem in CT imaging necessitates advanced reconstruction techniques.

Purpose of the Study:

• To develop a deep learning model for interpolating data in sparse-view sinograms.
• To address the trade-off between radiation exposure and image reconstruction quality in CT.
• To enhance the accuracy and efficiency of CT image reconstruction from limited projection data.

Main Methods:

• A deep learning model, based on super-resolution convolutional neural networks, was designed.
• The model takes sparse sinograms as input and outputs sinograms with interpolated data for additional views.
• Performance was evaluated by comparing mean-squared error (MSE) in reconstructed images.

Main Results:

• Reconstructions from model-interpolated sinograms demonstrated lower MSE than those from original sparse sinograms.
• The deep learning approach outperformed the bilinear image-resizing algorithm in reducing MSE.
• The model exhibits adaptability to various image sizes and efficient computational performance.

Conclusions:

• Deep learning offers a promising solution for the sparse-view problem in CT imaging.
• The developed model effectively interpolates sinogram data, leading to improved image reconstruction with reduced radiation exposure.
• This approach enhances CT diagnostic accuracy while optimizing time and memory efficiency.