Paired conditional generative adversarial network for highly accelerated liver 4D MRI.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• 4D MRI is crucial for image-guided liver radiotherapy, but acquiring high-resolution data is time-consuming.
• Accelerated MRI acquisition with sparse sampling often compromises image quality or increases reconstruction time.

Purpose of the Study:

• To propose and evaluate the Reconstruct Paired Conditional Generative Adversarial Network (Re-Con-GAN) for rapid and high-quality 4D liver MRI reconstruction.
• To reduce the reconstruction time of 4D MRI while preserving image fidelity for radiotherapy applications.

Main Methods:

• Developed Re-Con-GAN, a generative adversarial network, utilizing ResNet9, UNet, or reconstruction swin transformer generators with PatchGAN discriminator.
• Trained Re-Con-GAN on 4D liver MRI data (3D + time) from 48 patients, processing data as temporal slices (2D + time).
• Compared Re-Con-GAN against Compressed Sensing (CS) and UNet models, evaluating image quality using PSNR, SSIM, RMSE, and a liver gross tumor volume (GTV) localization task.

Main Results:

• Re-Con-GAN achieved comparable or superior PSNR, SSIM, and RMSE scores versus CS and UNet.
• Inference time for Re-Con-GAN was significantly faster (0.15s) compared to CS (120s) and comparable to UNet (0.16s).
• Re-Con-GAN improved the Dice score in GTV detection tasks on under-sampled images (80.98%) compared to UNet (79.88%).

Conclusions:

• Re-Con-GAN demonstrates efficient and high-quality 4D liver MR image reconstruction using adversarial training.
• The rapid and qualitative reconstruction capabilities of Re-Con-GAN can potentially enhance online adaptive MR-guided radiotherapy for liver cancer.