FastDIP: An effective approach for accelerating unsupervised low-count PET image reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Positron Emission Tomography (PET)

Background:

• Unsupervised deep learning enhances PET image quality without large datasets.
• Current methods are time-consuming due to patient-specific network training, limiting clinical use.
• There is a need for efficient unsupervised PET image reconstruction for real-time capabilities.

Purpose of the Study:

• To develop an efficient unsupervised learning framework for PET image reconstruction.
• To meet clinical requirements for real-time imaging.
• To reduce the computational burden of unsupervised PET image reconstruction.

Main Methods:

• Introduced FastDIP, an efficient unsupervised method for low-count PET image reconstruction.
• Employed a two-stage reconstruction: rapid coarse reconstruction using pixel-shuffle downsampling and detailed fine reconstruction.
• Utilized wavelet-denoised PET images as input and incorporated pre-training for accelerated convergence.

Main Results:

• FastDIP demonstrated superior image quality and significantly reduced training time (11% of DIP) on simulated 18F-AV45 datasets.
• On clinical 18F-FDG datasets, FastDIP achieved the lowest NMSE and highest SSIM in 2.2 minutes, outperforming existing methods.
• For clinical 68Ga-PSMA datasets, FastDIP yielded the highest CNR and SUVmax in 2.4 minutes, surpassing other approaches.

Conclusions:

• FastDIP offers an efficient solution for unsupervised low-count PET image reconstruction.
• The method substantially decreases network training time.
• FastDIP markedly enhances image restoration performance in PET imaging.