Unsupervised segmentation of dynamic pulmonary MRI using cross-modality adaptation with annotated CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate lung parenchyma segmentation in dynamic pulmonary MRI is crucial for diagnosis and treatment planning.
• Supervised deep learning methods require extensive annotated datasets, which are limited for pulmonary MRI.

Purpose of the Study:

• To develop an unsupervised segmentation framework for pulmonary MRI by leveraging annotated CT data.
• To overcome the challenge of scarce annotated pulmonary MRI datasets.

Main Methods:

• A novel framework involving a masked autoencoder for modality-invariant feature learning.
• Pretraining an initial segmenter with labeled CT data and temporal consistency loss on 4D MR images.
• Utilizing a select-and-refine pipeline to generate pseudolabels for training a final segmenter on MRI data.

Main Results:

• The method achieved high accuracy in lung parenchyma segmentation on 4D MR images, outperforming existing cross-modality techniques.
• Achieved Dice scores of ~97.75% and average surface distances of ~1.80 mm.
• Demonstrated robust performance across data from two different imaging centers.

Conclusions:

• The proposed method successfully transfers segmentation knowledge from CT to MRI, enabling accurate unsupervised segmentation of dynamic pulmonary MRI.
• This technique eliminates the need for MRI annotations, providing a practical and promising solution for clinical applications.