Denoising diffusion-based MRI to CT image translation enables automated spinal segmentation

Area of Science:

• Medical imaging
• Radiology
• Computational anatomy

Background:

• Automated segmentation of spinal magnetic resonance imaging (MRI) is crucial for scientific and clinical advancements.
• Accurate delineation of posterior spinal structures in MRI remains a significant challenge.

Purpose of the Study:

• To develop and evaluate methods for translating spinal MRI to computed tomography (CT) images.
• To enable the use of CT-based segmentation tools on MRI data.
• To generate whole spine segmentation from MRI, facilitating biomechanical modeling and clinical feature extraction.

Main Methods:

• Retrospective analysis of 263 paired CT/MR series.
• Comparison of 2D paired (Pix2Pix, DDIM) and unpaired (SynDiff) image-to-image translation techniques.
• Evaluation using peak signal-to-noise ratio and Dice Similarity Coefficients (DSC).
• Extension of 2D methods to 3D Pix2Pix and DDIM, utilizing landmark-based registration with a minimum of two landmarks per vertebra.

Main Results:

• Paired 2D translation methods and SynDiff showed comparable performance on paired data.
• DDIM in image mode yielded the highest image quality.
• Similar DSC values (0.77) were observed for SynDiff, Pix2Pix, and DDIM image mode.
• 3D translation significantly outperformed 2D approaches, achieving a DSC of 0.80.
• 3D translation resulted in anatomically accurate segmentations with higher spatial resolution than the original MRI.

Conclusions:

• Paired image-to-image translation from MRI to CT, requiring at least two landmarks per vertebra for registration, outperformed unpaired methods.
• 3D translation techniques produced anatomically correct segmentations, accurately capturing small structures like the spinous process.
• This translation enables the application of CT-based tools to MRI data, providing comprehensive spinal segmentation previously unavailable.