Adaptive rectification based adversarial network with spectrum constraint for high-quality PET image synthesis.

Area of Science:

• Nuclear medicine
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Positron emission tomography (PET) provides vital functional information for early brain disease diagnosis.
• Standard-dose PET (SPET) requires high radioactive tracer doses, raising health concerns.
• Low-dose PET (LPET) images suffer from increased noise and artifacts, hindering diagnosis.

Purpose of the Study:

• To develop a method for synthesizing high-quality SPET images from LPET images.
• To reduce radiation exposure in PET imaging while preserving diagnostic accuracy.
• To improve the performance of generative adversarial networks for medical image synthesis.

Main Methods:

• An adaptive rectification generative adversarial network with spectrum constraint (AR-GAN) was developed.
• An adaptive rectification network (AR-Net) estimates residuals to refine preliminary synthesized images.
• Spectral regularization enforces frequency domain consistency to preserve high-frequency details.

Main Results:

• AR-GAN effectively synthesizes SPET images from LPET data.
• The proposed method outperforms existing state-of-the-art image synthesis techniques.
• Validations on phantom and clinical datasets confirm the efficacy of AR-GAN.

Conclusions:

• AR-GAN offers a promising solution for high-quality PET imaging with reduced radiation dose.
• The technique enhances diagnostic capabilities for early brain disease detection.
• This approach has the potential to significantly impact clinical nuclear medicine practices.