Superiorization-inspired unrolled SART algorithm with U-Net generated perturbations for sparse-view and limited-angle...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unrolled algorithms are effective for challenging computed tomography (CT) image reconstruction tasks like low-dose, sparse-view, and limited-angle imaging.
• These algorithms integrate iterative reconstruction methods into neural networks, learning regularizers from data.
• Existing methods often focus on sparse-view or low-dose CT, with less exploration in limited-angle scenarios.

Purpose of the Study:

• Propose a novel unrolled algorithm for CT image reconstruction.
• Compare its performance against existing methods in sparse-view and limited-angle CT.
• Investigate the integration of deep learning with superiorization methodology for image reconstruction.

Main Methods:

• Developed a novel unrolled algorithm inspired by superiorization methodology.
• Utilized a modified U-net architecture to introduce learned perturbations during reconstruction.
• Trained the network end-to-end in a supervised manner.
• Evaluated performance on numerical experiments simulating sparse-view and limited-angle CT.

Main Results:

• The proposed algorithm achieved excellent results in both sparse-view and limited-angle CT scenarios.
• It outperformed several competing unrolled methods in limited-angle CT reconstruction.
• Performance was comparable or superior to existing methods in sparse-view CT.

Conclusions:

• The novel unrolled algorithm demonstrates significant potential for CT image reconstruction, particularly in challenging limited-angle cases.
• This work is a foundational step in applying deep learning within the superiorization methodology.
• The study highlights the impact of network architecture and the effectiveness of unrolled methods in limited-angle CT.