Projection Image Synthesis Using Adversarial Learning Based Spatial Transformer Network For Sparse Angle Sampling CT.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Sparse-angle X-ray tomography (CT) reconstruction suffers from significant noise and artifacts.
• Existing deep learning methods require extensive fully sampled data for training.
• Artifact removal while preserving image features is a key research challenge.

Purpose of the Study:

• To develop a novel data-driven approach for improving CT image reconstruction quality.
• To address the limitation of large data requirements in current deep learning models.
• To enhance the number of projection images using existing data for better reconstruction.

Main Methods:

• An adversarial learning-based spatial transformer network was developed for projection image synthesis.
• The method is inspired by video frame synthesis techniques.
• It focuses on pre-processing the sinogram by increasing projection data.

Main Results:

• The proposed model effectively synthesizes projection images, increasing their number.
• Simulation and experimental results demonstrate competitive performance against conventional algorithms.
• The approach shows promise in enhancing the quality of reconstructed CT images.

Conclusions:

• The developed spatial transformer network offers an effective solution for sparse-angle CT image reconstruction.
• This data-driven model can improve image quality by generating additional projection data.
• The method provides a viable alternative to conventional algorithms, especially when training data is limited.