Deep-Learning-Based Metal Artefact Reduction With Unsupervised Domain Adaptation Regularization for Practical CT...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep learning CT metal artifact reduction (MAR) methods struggle with domain gap, limiting generalization to real data.
• Unsupervised MAR methods trained on real data often yield unsatisfactory results due to indirect learning metrics.

Purpose of the Study:

• To propose UDAMAR, a novel MAR method utilizing unsupervised domain adaptation (UDA) to address the domain gap in CT MAR.
• To improve the performance and generalizability of CT MAR techniques by aligning feature spaces between simulated and practical data.

Main Methods:

• UDAMAR integrates a UDA regularization loss into a supervised image-domain MAR framework.
• An adversarial-based UDA approach is employed, focusing on low-level feature spaces to align simulated and practical artifact characteristics.
• The method learns MAR from labeled simulated data while extracting information from unlabeled real-world data.

Main Results:

• UDAMAR demonstrated superior performance over its supervised counterpart and two state-of-the-art unsupervised methods on clinical dental and torso datasets.
• Experiments on simulated data showed UDAMAR's efficacy, closely matching supervised methods and outperforming unsupervised ones.
• Ablation studies confirmed the robustness of UDAMAR concerning UDA regularization weight, feature layers, and data quantity.

Conclusions:

• UDAMAR offers a feasible and effective solution for practical CT MAR by mitigating domain discrepancies.
• The method's simple design and ease of implementation make it highly suitable for real-world clinical applications.
• UDAMAR successfully combines the strengths of supervised learning on simulated data with unsupervised adaptation to real data.