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

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Lung magnetic resonance imaging (MRI) offers radiation-free functional assessment but faces segmentation challenges due to low signal-to-noise ratio and weak contrast.
• Existing methods struggle with limited reliable annotations for supervised learning, hindering automated segmentation.
• Lung computed tomography (CT) provides clear structural delineation and abundant labels, making it a potential source for indirect supervision.

Purpose of the Study:

• To develop a label-free lung MRI segmentation method by addressing the annotation bottleneck.
• To enable accurate segmentation of lung MRI by translating it into a structurally consistent synthetic CT.
• To propose a misalignment-aware diffusion framework for cross-modality medical image translation.

Main Methods:

• A misalignment-aware diffusion framework for MRI-to-CT translation was developed.
• The framework incorporates three-channel diversity and elastic deformation to handle respiratory motion and acquisition discrepancies.
• Normalized mutual information was used as a conditioning signal to convey cross-modality alignment.

Main Results:

• The proposed method achieved a Dice score of 82.38% and a 95th-percentile Hausdorff distance of 33.32 mm for lung segmentation.
• This significantly improved boundary delineation compared to direct MRI input (55.94% and 147.59 mm).
• Results were comparable to direct CT segmentation (81.11% and 38.02 mm).

Conclusions:

• The study presents a practical approach for label-free lung MRI segmentation.
• Misalignment-aware conditioning is a principled strategy for cross-modality medical image translation.
• The developed framework offers a viable alternative for functional lung assessment using MRI.