Purpose: 4D-MRI has shown great potential for motion management in abdominal radiotherapy. However, 4D-MRI usually suffers from severe motion artifacts and acquisition noise. In this study, we proposed a 2.5D adversarial network to improve the image quality of 4D-MRI.
Methods: We designed our 2.5D model based on generative adversarial nets (GAN) to learn an accurate mapping from the 4D-MR images to corresponding high-quality (HQ) MR images. The dataset was obtained from twenty-one liver tumor patients undergoing radiotherapy. The 4D-MRI acquisition was performed using the TWIST volumetric interpolated breath-hold examination (TWIST-VIBE) MRI sequence. The HQ counterpart was T1w (breath-hold) 3D MRI. To make full use of neighboring information and avoid inter-slice artifacts, we designed our model to accept 2.5D input by combining the upper and lower adjacent slices into one image. Besides, We used U-Net in both the generator and discriminator. This structure makes it possible to independently predict the fidelity of every pixel. The loss function comprises three parts, a condition GAN objective, an L1-norm-based distance to lessen blurring artifacts, and a multi-scale structural similarity index (MS-SSIM) to restore high-frequency information. We chose the enhanced deep super-resolution network (EDSR) as the referencing method due to its state-of-the-art performance.
Results: The image quality of 4D-MRI in our method was largely improved; organ shapes showed better visibility with fewer artifacts and noise. In addition, the quantitative results indicated that our method outperformed EDSR with a lower MAE of 0.0721 ± 0.015(vs 0.0817), higher SSIM, and Laplacian of 0.7645 ± 0.042(vs 0.7018) and 207.1445 ± 19.869(vs 16.9437), respectively.
Conclusion: We successfully demonstrated a novel 2.5D adversarial network for 4D-MRI image quality enhancement. The model can suppress motion artifacts and noise, and restore texture details, eventually improving visibility. This technique enables refining the 4D-MR images and promises to improve motion management in abdominal radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: This research was partly supported by research grants of the General Research Fund (GRF 15102118), the University Grants Committee, Health, and Medical Research Fund (HMRF 06173276), the Food and Health Bureau, Hong Kong Special Administrative Regions.
Not Applicable / None Entered.