Unsupervised skull segmentation in MR images utilizing modality translation and super-resolution

Area of Science:

• Computational medicine
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Skull segmentation is crucial in computational medicine but challenging in MRI due to its focus on soft tissues.
• Existing CT-based skull segmentation methods pose radiation risks, necessitating MRI-based alternatives.
• Direct MRI skull segmentation is difficult because MRI resolution is often lower than CT.

Purpose of the Study:

• To develop an unsupervised method for skull segmentation directly from MRI data.
• To overcome the limitations of MRI in bone structure visualization and resolution.
• To provide a radiation-free alternative to CT-based skull segmentation.

Main Methods:

• Leveraging unsupervised modality translation techniques to convert MRI to CT-like images.
• Investigating various deep generative networks, including GANs, diffusion models, and contrastive learning.
• Employing a super-resolution approach to enhance the resolution of MRI-derived images.
• Developing a novel methodology for MRI-based skull segmentation.

Main Results:

• The proposed methodology achieves fast inference for volumetric data.
• Outperforms traditional supervised segmentation methods for skull segmentation.
• Surpasses a novel foundation medical segmentation model.
• Delivers superior results compared to other modality translation techniques.

Conclusions:

• Unsupervised modality translation offers a viable solution for MRI-based skull segmentation.
• The developed method provides a radiation-free and efficient alternative for clinical applications.
• This approach enhances the utility of MRI in tasks traditionally reliant on CT imaging.