AAPM CT metal artifact reduction grand challenge

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Processing

Background:

• Metal artifacts in CT scans significantly degrade image quality and diagnostic value.
• Existing methods lack standardized benchmarks for objective performance comparison.
• A universal benchmark is crucial for evaluating and advancing CT metal artifact reduction (MAR) techniques.

Purpose of the Study:

• Establish a clinically representative 2D MAR performance benchmark via the AAPM CT-MAR grand challenge.
• Facilitate objective comparison of diverse MAR algorithms.
• Provide a MAR training database and tools for future research and development.

Main Methods:

• Organized a grand challenge inviting participants to submit MAR algorithm results.
• Developed a hybrid simulation framework generating 14,000 CT training datasets with realistic metal objects.
• Evaluated 29 clinical datasets using eight image quality metrics and compared against a normalized metal artifact reduction (NMAR) reference.

Main Results:

• 106 teams registered, 26 completed; 92% utilized deep learning (DL) approaches (UNet, ResNet, GAN, diffusion models, transformers).
• 22% of teams combined sinogram- and image-domain processing.
• Over 70% of teams surpassed the NMAR baseline, demonstrating diverse and effective MAR methods.

Conclusions:

• The CT-MAR grand challenge successfully benchmarked state-of-the-art MAR algorithms.
• The hybrid data generation framework proved effective for creating large-scale, realistic MAR datasets.
• The publicly released MAR benchmark and training data will support ongoing MAR development and comparison.