Independent brain ¹⁸F-FDG PET attenuation correction using a deep learning approach with Generative Adversarial...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nuclear Medicine

Background:

• Attenuation correction (AC) in Positron Emission Tomography (PET) is crucial but challenging without transmission or CT data.
• External or morphological imaging data are typically used for AC, limiting standalone PET or PET/MRI applications.
• Machine learning offers potential for direct AC from non-attenuation-corrected PET (PET_NAC) data.

Purpose of the Study:

• To develop and validate a Generative Adversarial Network (GAN) based method for independent attenuation correction of brain 18F-FDG PET images using only PET_NAC data.
• To evaluate the accuracy and clinical feasibility of GAN-derived attenuation maps for PET image correction.

Main Methods:

• A deep learning GAN framework was trained on paired PET_NAC and CT head images from 50 patients.
• Pseudo-CT images were generated from PET_NAC data of 40 validation patients (20 technical, 20 clinical CNS disorder cases).
• Pseudo-CT images were used for subsequent attenuation correction, generating independently corrected PET data.

Main Results:

• Generated pseudo-CT images showed high resemblance to acquired CT images with minor anatomical differences.
• Quantitative analysis revealed <5% underestimation of Standardized Uptake Value (SUV) in all brain regions.
• Color-coded error maps indicated minimal average errors (±0%) with no regional bias.
• Image-based diagnoses in patients with neurological disorders were consistent between independently corrected and reference PET images.

Conclusions:

• Independent attenuation correction of brain 18F-FDG PET is feasible with high accuracy using the proposed deep learning GAN framework.
• The method is easy to implement and shows promise for standalone PET scanners or PET/MRI systems.
• Further clinical validation in larger cohorts is recommended to assess the method's full clinical performance.