Low-dimensional Manifold Constrained Disentanglement Network for Metal Artifact Reduction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks show promise for CT metal artifact reduction (MAR).
• Existing methods often rely on synthesized data, which may not accurately represent real-world artifacts.
• Current unsupervised methods struggle to recover fine structural details in artifact-affected images.

Purpose of the Study:

• To develop a novel method for CT metal artifact reduction (MAR) that overcomes limitations of existing approaches.
• To improve the recovery of small structural details in artifact-affected CT images.
• To leverage both paired and unpaired clinical data for enhanced MAR performance.

Main Methods:

• Propose a low-dimensional manifold (LDM) constrained disentanglement network (DN).
• Incorporate LDM constraints into the disentanglement network to address the ill-posed nature of MAR.
• Develop a hybrid optimization scheme utilizing both paired and unpaired data for improved MAR.

Main Results:

• The LDM-DN approach consistently improves MAR performance across various learning settings.
• The method demonstrates superior performance compared to competing methods on both synthesized and clinical datasets.
• LDM-DN effectively recovers small structural details that are often lost in other MAR techniques.

Conclusions:

• The proposed LDM-DN is an effective method for CT metal artifact reduction (MAR).
• Constraining the disentanglement network to a low-dimensional manifold significantly enhances artifact removal and detail recovery.
• The hybrid optimization scheme further boosts MAR performance, offering a robust solution for clinical applications.