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

Introduction:

Unsupervised deep learning methods can improve the image quality of positron emission tomography (PET) images without the need for large-scale datasets. However, these approaches typically require training a distinct network for each patient, making the reconstruction process extremely time-consuming and limiting their clinical applicability. In this paper, our research objective is to develop an efficient unsupervised learning framework for unsupervised PET image reconstruction, in order to fulfill the clinical requirement for real-time imaging capabilities.

Methods:

In this study, we present FastDIP, an efficient learning method for unsupervised low-count PET image reconstruction. FastDIP employs a two-stage reconstruction process, beginning with a rapid coarse reconstruction followed by a detailed fine reconstruction. The pixel-shuffle downsampling method is utilized to compress PET images and facilitate quick coarse reconstruction. Additionally, a wavelet-denoised PET image serves as input, replacing the traditional anatomical images. We also incorporate pre-training techniques to accelerate network convergence.

Results:

The efficacy of FastDIP was evaluated on simulated ¹⁸F-AV45 brain datasets, as well as clinical ¹⁸F-FDG brain and clinical ⁶⁸Ga-PSMA body datasets. FastDIP was compared to Deep Image Prior (DIP), Conditional Deep Image Prior (CDIP), Guided Deep Image Prior (GDIP), Self-supervised Pre-training DIP (SPDIP), Population Pre-training DIP (PPDIP) and various ablation methods. For the ¹⁸F-AV45 dataset, FastDIP achieved better image quality than DIP using only 11% training time of them across different count levels. In the ¹⁸F-FDG dataset, it achieved the lowest normalized mean square error and the highest structural similarity in just 2.2 min, outperforming DIP (10.7 min), CDIP (7.5 min), GDIP (9.8 min), SPDIP (166.7 min) and PPDIP (166.7 min). For the ⁶⁸Ga-PSMA dataset, FastDIP achieved the highest contrast-to-noise ratio and SUV_max in 2.4 min, surpassing DIP (10.7 min), CDIP (32.0 min) , GDIP (32.7 min), SPDIP (16.7 min) and PPDIP (37.5 min).

Conclusion:

FastDIP is an efficient approach for unsupervised low-count PET image reconstruction that significantly reduces the network training time and markedly enhances image restoration performance.