MR-based Attenuation Correction for Brain PET Using 3D Cycle-Consistent Adversarial Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiological Physics

Background:

• Quantitative accuracy in Positron Emission Tomography (PET) relies heavily on effective attenuation correction (AC).
• Direct derivation of attenuation coefficients from Magnetic Resonance (MR) images is not feasible in integrated PET/MR systems, necessitating alternative methods.
• Current AC methods often depend on registered Computed Tomography (CT) data or less accurate segmentation-based approaches.

Purpose of the Study:

• To develop a novel method for generating continuous attenuation correction (AC) maps from Dixon MR images for PET/MR systems.
• To eliminate the need for MR and CT image registration in the AC process.
• To evaluate the performance of the proposed method against existing AC techniques.

Main Methods:

• A 3D generative adversarial network (GAN) incorporating discriminative and cycle-consistency loss (Cycle-GAN) was developed.
• Modified 3D U-net architectures served as generative networks for pseudo-CT/MR image synthesis.
• 3D patch-based discriminative networks were utilized to differentiate generated images from true images.

Main Results:

• The Cycle-GAN framework successfully generated pseudo-CT images for AC.
• Quantitative evaluation using Dice coefficients and regional PET quantification demonstrated superior AC compared to Dixon segmentation and atlas methods.
• The proposed method achieved performance comparable to a Convolutional Neural Network (CNN) method that required registered MR and CT images.

Conclusions:

• The developed Cycle-GAN framework provides an effective solution for generating accurate attenuation correction maps from Dixon MR images in PET/MR systems.
• This method offers a viable alternative to CT-based AC, simplifying workflows and potentially improving quantitative accuracy.
• The approach shows significant promise for advancing quantitative PET/MR imaging without reliance on co-registered CT data.