Adaptive Sampling Method for Whole-Body Low-Dose Pet Reconstruction Based on Reconstruction Difficulty

Low-dose PET imaging has been widely studied to reduce patient radiation exposure and improve hospital resource utilization. However, reducing the dose often introduces noise and blurs the tissue contours, motivating the development of deep-learning methods for reconstructing low-dose PET into fulldose PET. While these models can explore information within given low-dose PET patches, they overlook the anatomical variations present in whole-body PET imaging. We observe that in random sampling for whole-body PET, the probability of sampling difficult-to-reconstruct regions is the same as that of easily reconstructable regions, resulting in an imbalance that hinders the overall performance of the trained networks. Therefore, we propose a difficulty-adaptive sampling method that increases the sampling probability of challenging regions, reinforcing the network’s fitting capability for the poorly reconstructed areas during training and ultimately improving the reconstruction quality. Additionally, we design a noise level map as a guidance input during network training, enabling the network to recognize the noise level of each voxel in the low-dose PET, thereby enhancing the overall performance. The proposed method achieved fourth place in the MICCAI 2023 Ultra-low Dose PET Imaging Challenge.