Purpose: Our goal is to synthesize high quality and accurate Amyloid PET images with only ultra-low-dose PET images as input by using Generative Adversarial Network (GAN).
Methods: 40 patients’ PET data was acquired with the injection of 330±30 MBq amyloid radiotracer 18F-florbetaben. The raw list mode PET data was reconstructed as the standard-dose ground truth and was randomly undersampled by a factor of 100 to reconstruct 1% low-dose PET scans. 32 volumes were used for training and the other 8 for testing. A 2D encoder-decoder network was implemented as the generator to synthesize a standard-dose image and a CNN based discriminator network was used to evaluate them. The two networks contested with each other to achieve accurate synthesis of standard-dose PET images with high visual quality from ultra-low-dose PET. Multi-slice input is used to reduce noise by providing the network with 2.5D information. Feature matching was applied to reduce the hallucinate structure. The image quality was evaluated by peak signal-to-noise ratio (PSNR), structural similarity (SSIM), mean square error (MSE), frequency domain blur measure (FBM) and edge blur measure (EBM) metrics.
Results: The synthesized PET images showed remarkable improvement on all quality metric compared with the low-dose images. Comparing with the state-of-art method, adversarial learning is essential to ensure image quality and mitigate the blurring in the generated image. Multi-slice input reduced random noise and feature matching suppressed the hallucinate structure.
Conclusion: Standard-dose Amyloid PET images can be synthesized from ultra-low-dose image by GAN. Applying adversarial learning, multi-slice input and feature matching technique are essential to ensure image quality.
|