A bias field correction workflow based on generative adversarial network for abdominal cancers treated with 0.35T MR‐LINAC

Purpose In this study, a bias field correction workflow was proposed to improve the flexibility and generalizability of the generative adversarial network (GAN) model for abdominal cancer patients treated with a 0.35T magnetic resonance imaging linear accelerator (MR‐LINAC) system. Methods Model training was performed using brain MR images acquired on a 3T diagnostic scanner, while model testing was performed using abdominal MR images obtained using a 0.35T MR‐LINAC system. The performance of the proposed workflow was first compared with the GAN model using root‐mean‐square error (RMSE), peak signal‐to‐noise ratio (PSNR), and structural similarity index measure (SSIM). To assess the impact of the workflow on image segmentation, it was also compared with the N4ITK algorithm. Segmentation was performed using the k‐means clustering algorithm with three clusters corresponding to air, fat, and soft tissue. Segmentation accuracy was then evaluated using the Dice similarity coefficient (DSC). Results The RMSE values were 30.59, 12.06, 10.37 for the bias field‐corrupted images (I IN ), GAN‐corrected images (I GAN ), and images corrected with the proposed workflow (I OUT ), respectively. Corresponding PSNR values were 42.34, 46.04, 47.04 dB, and SSIM values were 0.84, 0.96, 0.98. For segmentation accuracy, the mean DSC for air masks was 0.95, 0.97, and 0.97; for fat masks, 0.61, 0.71, and 0.74; and for soft tissue masks, 0.60, 0.68, and 0.69, corresponding to I IN , N4ITK‐corrected images (I N4ITK ), and I OUT , respectively Conclusion By effectively mitigating bias field artifacts, the proposed workflow has the potential to strengthen the clinical utility of MRI‐guided adaptive radiotherapy for abdominal cancers, ensuring safer and more accurate radiation delivery.