A deep learning-based whole-body solution for PET/MRI attenuation correction.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Deep convolutional neural networks show promise for PET attenuation correction (AC) in PET/MRI.
• Whole-body implementation is challenging due to anatomical variations and limited MRI field of view.

Purpose of the Study:

• Investigate a deep learning (DL) method to generate voxel-based synthetic CT (sCT) from Dixon MRI.
• Utilize sCT as a whole-body solution for PET AC in PET/MRI systems.

Main Methods:

• Trained a DL network using co-registered MRI and CT images from 15 patients with whole-body PET/CT and PET/MRI scans.
• Assessed AC map accuracy and PET image quantification using DL-based sCT (PET_sCT) versus an atlas-based method (PET_Atlas), with CT-based reconstruction (PET_CT) as reference.
• Performed voxel-wise and region-specific analyses (brain, lung, liver, spine, bone, aorta).

Main Results:

• DL-based sCT achieved a lower mean absolute error (62 HU) compared to the atlas-based method (109 HU).
• PET_sCT showed excellent correlation with PET_CT (R²=0.98) and lower PET quantification error (6.1%) than PET_Atlas (11.2%).
• PET_sCT demonstrated reduced average errors and variability across all analyzed regions.

Conclusions:

• A DL approach for whole-body PET AC in PET/MRI is feasible and yields more accurate results than conventional methods.
• Further validation with larger training datasets is recommended for enhanced robustness and accuracy.