Cross-Scanner Low-Dose Brain-PET Image Noise Reduction With Self-Ensembling

Deep learning models have shown great potential in reducing low-dose (LD) positron emission tomography (PET) image noise by estimating full-dose (FD) images from the corresponding LD images. Those models, however, when trained on paired LD-FD PET images from a source scanner, fail to generalize well when applied to LD PET images from a target scanner, due to a phenomenon called "domain drift." In this study, we present a method for cross-scanner LD PET image noise reduction. This is done via a self-ensembling framework using a limited number of paired LD-FD PET images and a large number of LD PET images from the target scanner. The self-ensembling framework leverages the paired 2-D slices from both scanners to learn a regression model. It additionally incorporates a consistency loss on the LD PET images from the target scanner to enhance the model's generalization capability. We conduct experiments on three datasets, respectively, acquired from three different scanners, including a GE Discovery MI (DMI) scanner, a Siemens Biograph Vision 450 (Vision) scanner, and a UI uMI 780 (uMI) scanner. Results from our comprehensive experiments demonstrate the generalization capability of our method.