An unsupervised sparse-view CT reconstruction framework using combination of iterative deep image prior and ADMM

Area of Science:

• Medical Imaging
• Computational Imaging
• Deep Learning

Background:

• Sparse-view and low-dose computed tomography (CT) reduce radiation exposure but introduce artifacts and noise.
• Supervised deep learning (DL) methods excel in CT reconstruction but require extensive paired datasets.
• Existing methods face limitations due to data dependency and image quality degradation.

Purpose of the Study:

• Introduce ADMM-DRP, an unsupervised deep learning (DL) framework for CT reconstruction.
• Integrate an untrained neural network with the alternating direction method of multipliers (ADMM) iterative algorithm.
• Address limitations of supervised DL by reducing reliance on large training datasets.

Main Methods:

• Employ an untrained neural network as an image generator to optimize data inconsistency in the Radon domain.
• Utilize ADMM with total variation (TV) regularization to continuously update neural network inputs.
• Prevent overfitting common in traditional deep image prior (DIP) methods.

Main Results:

• Demonstrate superior performance of ADMM-DRP over conventional supervised and iterative methods in sparse-view and low-dose CT reconstruction.
• Achieve improved metric and visual quality in reconstructed CT images.
• Validate the framework's effectiveness through experimental evaluations.

Conclusions:

• ADMM-DRP significantly reduces the need for extensive training data in CT reconstruction.
• The framework achieves excellent performance in challenging sparse-view and low-dose CT scenarios.
• ADMM-DRP shows substantial potential for advancing medical imaging reconstruction techniques.