CNN-Based Projected Gradient Descent for Consistent CT Image Reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Imaging

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for solving inverse problems in medical imaging.
• Existing CNN-based methods often lack measurement consistency feedback, a crucial component of iterative algorithms.
• Projected Gradient Descent (PGD) is a common iterative reconstruction technique.

Purpose of the Study:

• To introduce a novel image reconstruction method integrating CNNs into the PGD framework.
• To address the lack of measurement consistency in purely data-driven CNN reconstruction approaches.
• To improve the performance of sparse-view computed tomography (CT) reconstruction.

Main Methods:

• Replaced the projector in PGD with a CNN.
• Developed a relaxed PGD algorithm incorporating a CNN for projection and gradient descent for measurement consistency.
• Proposed a training scheme for the CNN to function as a projector within the iterative process.

Main Results:

• The proposed algorithm guarantees convergence, potentially to a local minimum for non-convex problems.
• Demonstrated improved sparse-view CT reconstruction compared to Total Variation, dictionary learning, and a state-of-the-art deep learning method.
• The CNN effectively acts as a projector, guiding the reconstruction towards desired image properties.

Conclusions:

• The CNN-integrated PGD method offers a robust and effective solution for inverse problems in medical imaging.
• This hybrid approach combines the strengths of iterative methods (measurement consistency) and deep learning (image prior enforcement).
• The technique shows significant promise for enhancing image quality in low-dose or limited-view CT scans.