Sparsier2Sparse: Self-supervised convolutional neural network-based streak artifacts reduction in sparse-view CT...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Sparse-view computed tomography (CT) reduces scan time and radiation dose but introduces streak artifacts.
• Existing fully-supervised methods require paired full-view and sparse-view data, which is clinically infeasible.
• Novel self-supervised learning approaches are needed for artifact reduction in sparse-view CT.

Purpose of the Study:

• To propose a novel self-supervised convolutional neural network (CNN) method for streak artifact reduction in sparse-view CT images.
• To develop a method that does not require full-view CT data for training.
• To improve the diagnostic quality of CT images reconstructed from limited projection data.

Main Methods:

• A self-supervised CNN was trained using only sparse-view CT data.
• Prior images were generated by iteratively applying the trained network to estimate streak artifacts.
• Estimated artifacts were subtracted from sparse-view images to obtain artifact-reduced results.

Main Results:

• The proposed method was validated on XCAT and the AAPM Low-Dose CT Grand Challenge datasets.
• Visual inspection and modulation transfer function (MTF) analysis demonstrated effective preservation of anatomical structures.
• The method achieved higher image resolution compared to existing artifact reduction techniques.

Conclusions:

• A novel framework for streak artifact reduction using only sparse-view CT data was developed.
• The self-supervised method achieved superior performance in preserving fine details without full-view data.
• This framework overcomes dataset limitations of fully-supervised methods and has potential for medical imaging applications.