Generating Synthetic T2*-Weighted Gradient Echo Images of the Knee with an Open-source Deep Learning Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Deep Learning for MRI

Background:

• Routine knee MRI protocols often omit T2*-weighted (T2*W) gradient echo sequences.
• T2*W images are valuable for assessing cartilage, synovial hemosiderin deposition, chondrocalcinosis, and pediatric physeal evaluation.
• Current MRI protocols may not fully capture all critical knee pathologies.

Purpose of the Study:

• To develop an open-source deep learning model for generating synthetic T2*W knee MRI images.
• To utilize fat-suppressed intermediate-weighted images as input for synthetic T2*W image generation.
• To provide a freely accessible tool for enhancing knee MRI diagnostics.

Main Methods:

• A cycleGAN deep learning model was trained on over 12,000 sagittal knee MR images.
• The model was tested on an independent dataset of nearly 3,000 images.
• Diagnostic interchangeability and image quality (NRMSE, SSIM, PSNR) were rigorously assessed.

Main Results:

• The synthetic T2*W images demonstrated high diagnostic interchangeability (ICC > 0.95).
• Image quality metrics (median NRMSE=0.5, PSNR=17.4, SSIM=0.5) indicate good performance.
• Identified artifacts had minimal to no impact on diagnostic value.

Conclusions:

• The developed open-source GAN model effectively generates high-quality synthetic T2*W knee MRI images.
• The synthetic images possess significant diagnostic value, comparable to conventional T2*W scans.
• This tool can potentially improve routine knee MRI examinations without requiring additional scan time.