TAI-GAN: A Temporally and Anatomically Informed Generative Adversarial Network for early-to-late frame conversion in...

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Dynamic cardiac positron emission tomography (PET) using rubidium-82 (82Rb) is crucial for myocardial perfusion imaging.
• Inter-frame motion significantly impacts myocardial blood flow (MBF) quantification and coronary artery disease diagnosis.
• Rapid tracer kinetics in early frames present challenges for traditional motion correction methods.

Purpose of the Study:

• To develop a novel method for correcting inter-frame motion in dynamic cardiac 82Rb PET imaging.
• To improve the accuracy of MBF quantification and diagnostic performance in coronary artery disease detection.

Main Methods:

• Proposed a Temporally and Anatomically Informed Generative Adversarial Network (TAI-GAN) for converting early PET frames.
• TAI-GAN uses an all-to-one mapping with temporal information and anatomical segmentation masks.
• Employed a feature-wise linear modulation layer for encoding channel-wise parameters.

Main Results:

• TAI-GAN successfully converted early frames to achieve high image quality comparable to reference frames.
• Motion estimation accuracy was improved after TAI-GAN conversion.
• MBF quantification accuracy was enhanced using both conventional and deep learning-based motion correction methods post-conversion.

Conclusions:

• TAI-GAN effectively addresses the challenge of inter-frame motion in dynamic cardiac PET imaging.
• The proposed method enhances the reliability of MBF quantification and aids in diagnosing coronary artery disease.
• TAI-GAN offers a promising solution for improving the clinical utility of cardiac PET imaging.