PIDNET: Polar Transformation Based Implicit Disentanglement Network for Truncation Artifacts.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Truncation artifacts in Computed Tomography (CT) imaging are a significant challenge, distorting CT values and impacting clinical diagnoses.
• Supervised deep learning models offer solutions but require extensive paired data, limiting their practical use.
• Existing methods struggle to effectively address the ill-posed nature of the interior problem in CT.

Purpose of the Study:

• To develop a simple, efficient, and unsupervised method for removing truncation artifacts in CT imaging.
• To overcome the data dependency of supervised deep learning models for artifact correction.
• To enhance the accuracy and reliability of CT image analysis by mitigating truncation artifacts.

Main Methods:

• A Cycle-GAN based unsupervised framework was employed, incorporating an implicit disentanglement strategy.
• Truncation artifact features were separated from content information for use as complementary constraints.
• Polar transformation and a novel constraint for artifact features were integrated to improve artifact removal.
• Simulated paired data was generated using separated artifact features to train a sub-network.

Main Results:

• The proposed unsupervised network significantly outperformed the traditional Cycle-GAN model.
• Comparable visual results were achieved when compared to state-of-the-art supervised models.
• Quantitative evaluation metrics closely aligned with those of supervised methods.

Conclusions:

• The developed unsupervised method effectively removes truncation artifacts in CT imaging without requiring paired data.
• This approach offers a practical and efficient alternative to supervised methods, broadening the applicability of deep learning in CT artifact correction.
• The implicit disentanglement strategy and specialized constraints enhance the performance of unsupervised artifact removal techniques.