Denoising of ultra-low-dose 15O positron emission tomography images using deep image prior with anatomical information extracted through magnetic resonance segmentation.

PURPOSE: Recent deep-learning methods can recover standard-dose PET images from low-dose images. However, these methods require a large amount of data. We proposed a novel unsupervised learning approach using conditional deep image prior (DIP) with tissue probability maps extracted by magnetic resonance (MR) segmentation (DIPseg), to improve denoising performance by more explicit anatomical guidance. Further, we retrospectively applied the proposed DIPseg method to recover standard-dose C15O2 PET images from ultra-low-dose images. METHODS: Ultra-low-dose C15O2 PET data for 10 patients with cerebrovascular steno-occlusive disease were generated by decimating the list-mode data by 1/128. The reconstructed ultra-low-dose images were denoised using DIPseg, DIP with MR T1-weighted images (DIPmr), image-guided filtering (IGF), and Gaussian filtering (GF). Biases from the full-dose PET images and the coefficient of variation (CoV) for the ultra-low-dose datasets were calculated to evaluate quantitative accuracy and image noise, respectively. To validate the ability to improve the quantification quality, cerebral blood flow (CBF) maps were estimated using the autoradiographic method. RESULTS: DIPseg achieved a significantly lower CoV (∼10%) than the other methods, while maintaining a low bias (∼3%) compared to the full-dose images. Unlike DIPmr, DIPseg suppressed the increase in image noise during the iterations. CBF quantified using DIPseg showed high similarity to that from the full-dose images. CONCLUSION: DIPseg can effectively recover the quality of standard-dose PET images from ultra-low-dose data, enabling accurate CBF quantification. The proposed DIPseg method has the potential to reduce radiation exposure in PET imaging, while maintaining diagnostic quality.