Unsupervised CT Metal Artifact Learning Using Attention-Guided β-CycleGAN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Metal artifact reduction (MAR) is critical in computed tomography (CT) imaging.
• Supervised deep learning methods for MAR require difficult-to-obtain matched artifact-free and corrupted image pairs.
• Existing unsupervised MAR methods are often complex and unsuitable for large clinical images.

Purpose of the Study:

• To develop a simple and effective unsupervised learning method for MAR.
• To address limitations of current supervised and unsupervised MAR techniques.
• To improve the quality of CT images corrupted by metal artifacts.

Main Methods:

• Proposed a novel β-cycleGAN architecture based on optimal transport theory for feature disentanglement.
• Integrated the convolutional block attention module (CBAM) into the generator to enhance artifact focus.
• Employed an unsupervised learning strategy to avoid the need for paired training data.

Main Results:

• Achieved significant improvement in metal artifact reduction compared to existing methods.
• Demonstrated effective preservation of detailed image textures.
• The proposed method is computationally efficient and handles large clinical images.

Conclusions:

• The novel unsupervised β-cycleGAN with CBAM offers a simple yet effective solution for MAR in CT.
• This method overcomes the data acquisition challenges of supervised learning.
• The approach holds promise for enhancing diagnostic accuracy in CT imaging.