Dixon-VIBE Deep Learning (DIVIDE) Pseudo-CT Synthesis for Pelvis PET/MR Attenuation Correction

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Nuclear Medicine

Background:

• Whole-body attenuation correction (AC) in combined PET/MR scanners remains a significant challenge.
• Accurate AC is crucial for reliable quantitative analysis of PET data.

Purpose of the Study:

• To introduce Dixon-VIBE Deep Learning (DIVIDE), a novel deep learning network for synthesizing pelvis pseudo-CT maps.
• To enable accurate AC in combined PET/MR scanners using only standard Dixon-VIBE MR images.

Main Methods:

• A deep learning network was developed to map four 2D Dixon MR images (water, fat, in-phase, out-of-phase) to corresponding 2D CT images.
• The network utilized transposed convolutions for up-sampling and whole 2D slices for context, pre-trained with brain images.
• Evaluation involved comparing SUV quantification from PET_DIVIDE, PET_Dixon, and PET_CT across 28 datasets from 19 patients.

Main Results:

• DIVIDE achieved mean relative changes below 2% compared to CT AC (except bone <4%), significantly outperforming the Dixon method.
• Excellent voxel-by-voxel correlation was observed between PET_CT and PET_DIVIDE (R² = 0.9998).
• Bland-Altman analysis showed lower bias and variability for PET_DIVIDE compared to PET_Dixon, with significant improvements in synthetic lesions (femur, spine).

Conclusions:

• The DIVIDE method accurately synthesizes pelvis pseudo-CT scans from Dixon-VIBE images for precise AC in PET/MR scanners.
• The rapid synthesis capability makes DIVIDE suitable for routine clinical applications and retrospective data processing.