Radiologist-in-the-Loop Self-Training for Generalizable CT Metal Artifact Reduction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Metal artifacts in computed tomography (CT) images degrade quality and hinder diagnosis.
• Supervised metal artifact reduction (MAR) methods fail on real images due to domain gaps.
• Current semi-supervised MAR methods have limited pseudo ground-truth quality and quantity.

Purpose of the Study:

• To develop a novel radiologist-in-the-loop self-training framework for MAR.
• To enhance the generalization of MAR models on real clinical CT images.
• To address limitations of existing MAR methods, including domain gap and confirmation bias.

Main Methods:

• Proposed a radiologist-in-the-loop self-training framework (RISE-MAR).
• Integrated radiologists' feedback to improve pseudo ground-truth quality and quantity.
• Introduced a clinical quality assessor model for pseudo ground-truth selection.
• Implemented iterative pseudo ground-truth generation for dataset expansion.

Main Results:

• RISE-MAR demonstrated superior generalization performance on clinical datasets.
• The framework effectively improved MAR model performance compared to state-of-the-art methods.
• Enhanced image quality and diagnostic accuracy in real-world CT scans.

Conclusions:

• RISE-MAR advances practical application of MAR models.
• The radiologist-in-the-loop approach mitigates domain gap issues.
• This framework offers a promising solution for improving CT image analysis.