VQPET: Leveraging Vector-Quantized Codebook Prior for PET Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Positron Emission Tomography (PET) image synthesis is challenging due to information loss and mapping uncertainty in low-quality inputs.
• Existing auxiliary priors often lack accuracy and expressiveness, limiting their effectiveness in PET image synthesis.
• Vector-quantized (VQ) codebook priors offer a promising solution by learning discrete latent features from high-quality (HQ) images.

Purpose of the Study:

• To introduce VQPET, a novel two-stage framework for PET image synthesis.
• To leverage a VQ codebook prior for enhanced guidance and reduced uncertainty in PET image synthesis.
• To improve the accuracy and expressiveness of auxiliary priors in PET image generation.

Main Methods:

• A two-stage framework, VQPET, was developed for PET image synthesis.
• Stage one involves pretraining a VQGAN on a large-scale HQ PET dataset to create a VQ codebook prior.
• Stage two utilizes a codebook-prior-guided network (CPGNet) that integrates the VQ codebook prior at multiple decoding levels.

Main Results:

• The VQ codebook prior, derived from HQ PET data, provides accurate and expressive guidance.
• CPGNet effectively utilizes the VQ codebook prior for generating realistic PET images.
• VQPET demonstrates substantial improvements in PET image synthesis quality and generalizability across different networks.

Conclusions:

• VQPET innovatively employs a VQ codebook prior for superior PET image synthesis.
• The framework effectively transfers prior knowledge from large-scale HQ datasets.
• VQPET offers a generalizable solution applicable to various encoder-decoder architectures in medical imaging.