Misalignment-Aware MRI-to-CT Synthesis for Lung Segmentation on MRI

Lung MRI is increasingly utilized for its radiation-free nature and ability to reflect lung function, but weak lung signals make accurate segmentation difficult. In addition, the lack of labeled datasets, as manual labeling is labor-intensive, poses a challenge for lung MRI segmentation. Noting that pre-trained lung CT segmentation models are widely available, we propose a novel framework that applies MRI-to-CT translation using a diffusion model to address the issue. The synthesized CT can be easily applied to the pre-trained lung CT segmentation model. Our approach also resolves misalignment issues caused by differences in acquisition principles and breathing techniques (free breathing vs. breath-hold) between MRI and CT, while effectively capturing MRI's structural details and CT's lung-specific information to enhance segmentation accuracy. The proposed method surpasses existing techniques in both quantitative metrics, such as Dice coefficient and Hausdorff distance, and qualitative evaluations, particularly in difficult-to-segment MRI slices. This research sets a new standard for lung MRI segmentation, reducing the need for MRI-specific labels and paving the way for future advancements in the field. Our code is available at https://github.com/Nejung-Rue/MR2CTforLungSeg.