Label-Free Lung MRI Segmentation via Misalignment-Aware Diffusion Translation.

Lung magnetic resonance imaging (MRI) is an attractive radiation-free modality for functional lung assessment, yet automated segmentation remains challenging due to low signal-to-noise ratio and weak boundary contrast, severely limiting the availability of reliable annotations for supervised learning. Lung computed tomography (CT), by contrast, provides clear structural delineation and abundant labels. Leveraging CT labels as indirect supervision for lung MRI segmentation is a natural choice. However, respiratory-phase differences and acquisition discrepancies often introduce residual MRI-CT misalignment, making direct label transfer unreliable even with registration. We propose a misalignment-aware diffusion framework for MRI-to-CT translation that directly addresses the annotation bottleneck by enabling label-free lung MRI segmentation through generating a structurally consistent synthetic CT from MRI and applying a well-established CT-trained segmentation model. Our framework includes misalignment-aware designs, such as three-channel diversity to reflect the physical characteristics of CT and MRI and elastic deformation to handle respiratory motion and acquisition differences. It also incorporates normalized mutual information as a numerical conditioning signal that conveys the degree of cross-modality alignment. On the test set, the proposed method achieved a Dice score of 82.38% and a 95th-percentile Hausdorff distance of 33.32 mm for both-lung segmentation, indicating more reliable segmentation with improved boundary delineation compared with direct MRI input (55.94% and 147.59 mm). Our results were comparable to those obtained with direct CT (81.11% and 38.02 mm). Overall, this work offers a practical route to label-free lung MRI segmentation and motivates misalignment-aware conditioning as a principled strategy for cross-modality medical image translation. The code is at https://github.com/Nejung-Rue/NMISynCT-LungSeg .