Self-supervised U-transformer network with mask reconstruction for metal artifact reduction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Metal artifacts in computed tomography (CT) scans obscure crucial human tissue information, hindering accurate disease diagnosis.
• Deep learning approaches for metal artifact reduction (MAR) face challenges due to the lack of paired real-world training datasets.
• Existing methods using synthetic data often exhibit poor generalization to real metal artifact CT images.

Purpose of the Study:

• To develop a self-supervised U-shaped transformer network for enhanced generalizability in metal artifact reduction (MAR).
• To improve the accuracy of disease diagnosis by effectively reducing metal artifacts in CT images.
• To leverage both unlabeled real-artifact and labeled synthetic-artifact CT images for robust model training.

Main Methods:

• A self-supervised U-shaped transformer network incorporating a mask reconstruction pre-text task was designed.
• The pre-text task involved reconstructing masked CT images to learn artifact and tissue structures.
• A downstream task fine-tuned the network for MAR using labeled data, utilizing Transformer's long-range feature extraction and a MAR bottleneck for cross-channel attention.

Main Results:

• The proposed framework demonstrated strong generalization for MAR, preserving tissue details while suppressing artifacts.
• Achieved a peak signal-to-noise ratio of 43.86 dB and a structural similarity index of 0.9863.
• Improved segmentation performance with an 11.70% increase in Dice coefficient and 9.51% in mean intersection over union for MAR images.

Conclusions:

• The self-supervised approach effectively enhances model generalizability for metal artifact reduction (MAR) in CT imaging.
• Combining unlabeled real and labeled synthetic data is crucial for improving model performance on real-world artifact data.
• The developed method offers an efficient solution for improving diagnostic accuracy in CT scans affected by metal artifacts.