Synthetic Inflammation Imaging with PatchGAN Deep Learning Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Rheumatology

Background:

• Gadolinium-enhanced MRI is vital for diagnosing rheumatoid arthritis (RA) but carries safety concerns.
• Synthetic generation of post-contrast MRI from non-contrast scans offers clinical utility for RA.
• AI models for musculoskeletal MRI synthesis are underexplored, with limited trust-building efforts.

Purpose of the Study:

• To develop and evaluate AI algorithms for synthesizing post-gadolinium contrast wrist MR images from pre-contrast scans in RA patients.
• To compare the performance of UNet and PatchGAN models for synthetic MRI generation in RA.
• To assess model interpretability using occlusion and uncertainty maps.

Main Methods:

• Trained UNet and PatchGAN models on 27 RA patients' wrist MRI data.
• Utilized anomaly-weighted L1 loss and global GAN loss for model training.
• Generated occlusion and uncertainty maps to analyze model predictions and confidence.

Main Results:

• PatchGAN demonstrated superior performance and confidence in synthesizing post-contrast images within synovial joints compared to UNet.
• UNet showed lower normalized root mean square error (nRMSE) in overall volume and wrist, but higher in synovial joints.
• Occlusion and uncertainty maps indicated significant contributions of synovial joints to predictions, with PatchGAN showing higher confidence.

Conclusions:

• AI-driven synthetic MRI generation shows promise for RA imaging, potentially reducing gadolinium use.
• PatchGAN models offer a more robust and confident approach for synthesizing contrast-enhanced images in critical RA joint areas.
• Further development of interpretable AI models is crucial for clinical trust and adoption in musculoskeletal imaging.