Deep learning for whole-body medical image generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep convolutional networks and generative adversarial networks (GANs) excel at image transformation.
• These AI techniques are increasingly applied to medical imaging for cross-domain translation.
• Unsupervised learning methods are advancing medical image synthesis.

Purpose of the Study:

• To investigate deep learning for whole-body MR-to-CT image transformation.
• To assess the quality and anatomical accuracy of synthetic CT for PET attenuation correction (AC).
• To compare GAN-based AC with current MR-based AC (MR-AC) methods.

Main Methods:

• Utilized whole-body MR data from PET/MR systems to generate synthetic CT volumes.
• Employed a generative adversarial network (GAN) system for MR-to-CT volumetric transformation.
• Validated performance using matched PET/MR and PET/CT patient scans.

Main Results:

• Generated high-quality, anatomically accurate synthetic CT images.
• Synthetic CT demonstrated higher correlation in mu maps compared to segmented Dixon MR-AC.
• Synthetic CT-based AC showed improved quantification accuracy in specific regions over MR-AC.

Conclusions:

• Demonstrated the feasibility of GAN-based synthetic CT for PET/MR attenuation correction.
• The technique holds potential for improving PET/MR AC and enabling generalized inter-modality image transformation.
• Further research may be needed to ensure medical integrity with unsupervised deep learning image generation.