Adaptive Weighting Based Metal Artifact Reduction in CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology

Background:

• Current deep learning methods for metal artifact reduction (MAR) in computed tomography (CT) often use a single Hounsfield unit (HU) window, limiting effectiveness in clinical practice where multiple windows are used.
• Existing multi-window MAR approaches apply equal weighting, hindering learning flexibility and model generalization.

Purpose of the Study:

• To develop an adaptive weighting algorithm (AdaW) for multi-window MAR to enhance deep learning model generalizability and applicability.
• To address the limitations of fixed single-window preprocessing and equal weighting in existing MAR techniques.

Main Methods:

• Formulated the multi-window MAR task as a bi-level optimization problem.
• Derived an adaptive weighting optimization algorithm (AdaW) using a learning-to-learn paradigm with training and validation sets.
• AdaW enables automatic weighting of MAR learning across different windows, applicable to various deep MAR network backbones.

Main Results:

• AdaW demonstrated improved generalization performance and applicability across different network backbones and five diverse datasets.
• Experimental comparisons validated the effectiveness of AdaW in enhancing multi-window MAR.

Conclusions:

• AdaW offers a flexible and effective solution for multi-window MAR, improving upon existing methods.
• The proposed adaptive weighting strategy enhances the robustness and clinical utility of deep learning models for CT MAR.