Deep learning-based pseudo-CT synthesis from zero echo time MR sequences of the pelvis.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Conventional CT scans involve radiation exposure, posing risks for repeated imaging.
• MRI offers excellent soft-tissue contrast but lacks the bone detail of CT.
• Developing radiation-free imaging alternatives is crucial for patient safety and frequent monitoring.

Purpose of the Study:

• To synthesize pseudo-CT (pCT) images from pelvic ZTE MRI sequences.
• To evaluate the accuracy and quality of synthesized pCT images compared to conventional CT.
• To explore the potential of pCT in reducing radiation exposure in pelvic imaging.

Main Methods:

• Prospective CT and ZTE MRI scans of 91 patients.
• Exclusion of 11 datasets due to implants or B1 inhomogeneity.
• Training a deep learning (DL) model on 60 datasets for pCT synthesis from ZTE sequences.
• Validation of the DL model on a separate cohort of 20 cases.
• Qualitative and quantitative assessment of CT and pCT images by two independent readers.

Main Results:

• pCT images demonstrated good to perfect qualitative ratings.
• High intermodality agreement (ICC=0.88) between CT and pCT.
• Excellent interreader agreement for pCT (ICC=0.91).
• No significant differences in most geometrical measurements, except for transverse pelvic diameter and lateral center-edge angle.
• Comparable image quality and tissue differentiation between CT and pCT, with similar contrast-to-noise ratios (CNRs).

Conclusions:

• DL-based pCT synthesis from ZTE MRI is feasible and accurate for pelvic imaging.
• pCT images provide high bone depiction quality and reliable geometrical measurements.
• pCT offers a radiation-free alternative to CT for evaluating bone, with broad applications in degenerative diseases and surgical planning.