ADN: Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep learning methods for CT metal artifact reduction (MAR) require synthesized data.
• Synthesized data may not accurately reflect real-world CT imaging physics, leading to poor generalization in clinical settings.

Purpose of the Study:

• To develop the first unsupervised learning approach for CT metal artifact reduction (MAR).
• To address the generalization limitations of supervised MAR methods in clinical applications.

Main Methods:

• Introduction of a novel artifact disentanglement network.
• Disentangling metal artifacts from CT images in the latent space.
• Utilizing specialized loss functions for artifact reduction, transfer, and self-reconstruction without synthesized data.

Main Results:

• The proposed unsupervised method significantly outperforms existing unsupervised models for natural image-to-image translation on MAR tasks.
• It achieves performance comparable to supervised MAR models on synthesized datasets.
• Demonstrates superior generalization ability on clinical datasets compared to supervised methods.

Conclusions:

• The novel unsupervised learning approach effectively performs CT metal artifact reduction.
• This method overcomes the limitations of supervised approaches by avoiding reliance on synthesized data.
• The technique shows promising generalization capabilities for real-world clinical CT imaging.