End-to-end Triple-domain PET Enhancement: A Hybrid Denoising-and-reconstruction Framework for Reconstructing Standard-dose PET Images from Low-dose PET Sinograms

TL;DR

TriPLET addresses reconstructing SPET from LPET by leveraging a triple-domain representation (projection, wavelet, and image domains) within an end-to-end hybrid denoising-and-reconstruction pipeline. It combines a Transformer-assisted sinogram denoising network, a WT-based wavelet-domain reconstruction network, and a paired image-domain adversarial discriminator, all guided by domain-specific losses and GradNorm balancing. Experimental results on real chest–abdomen PET data show TriPLET achieving leading PSNR, SSIM, and rRMSE, with external validation demonstrating robustness across datasets and scanners. The approach offers a practical pathway to reduce radiation exposure while preserving diagnostic image quality in PET imaging.

Abstract

As a sensitive functional imaging technique, positron emission tomography (PET) plays a critical role in early disease diagnosis. However, obtaining a high-quality PET image requires injecting a sufficient dose (standard dose) of radionuclides into the body, which inevitably poses radiation hazards to patients. To mitigate radiation hazards, the reconstruction of standard-dose PET (SPET) from low-dose PET (LPET) is desired. According to imaging theory, PET reconstruction process involves multiple domains (e.g., projection domain and image domain), and a significant portion of the difference between SPET and LPET arises from variations in the noise levels introduced during the sampling of raw data as sinograms. In light of these two facts, we propose an end-to-end TriPle-domain LPET EnhancemenT (TriPLET) framework, by leveraging the advantages of a hybrid denoising-and-reconstruction process and a triple-domain representation (i.e., sinograms, frequency spectrum maps, and images) to reconstruct SPET images from LPET sinograms. Specifically, TriPLET consists of three sequentially coupled components including 1) a Transformer-assisted denoising network that denoises the inputted LPET sinograms in the projection domain, 2) a discrete-wavelet-transform-based reconstruction network that further reconstructs SPET from LPET in the wavelet domain, and 3) a pair-based adversarial network that evaluates the reconstructed SPET images in the image domain. Extensive experiments on the real PET dataset demonstrate that our proposed TriPLET can reconstruct SPET images with the highest similarity and signal-to-noise ratio to real data, compared with state-of-the-art methods.

Figures (6)

• Figure 1: Representation of PET data in three domains: (a) projection domain, (b) image domain, and (c) wavelet domain, with the SPET and LPET shown on the left and right of each subfigure, respectively. Radon transform and discrete wavelet transform can convert PET data between images and sinograms, and also between images and frequency spectrum maps, respectively. After applying discrete wavelet transform, a 3D PET image can be decomposed into eight frequency spectrum maps of different sub-bands.
• Figure 2: Our proposed TriPLET involves three domains, i.e., the projection, wavelet, and image domains. PET data have different representations in each domain, such as the LPET sinogram $S_{Low}$ and the denoised sinogram $S_{Den}$ in the projection domain, the denoised frequency spectrum map $F_{Den}$ and the predicted SPET map $\hat{F}_{Std}$ in the wavelet domain, and the LPET image $I_{Low}$, the predicted SPET image $\hat{I}_{Std}$, and the actual SPET image $I_{Std}$ in the image domain. We have also designed specific loss functions $\mathcal{L}_{P}$, $\mathcal{L}_{F}$, and $\mathcal{L}_{I}$ to supervise the prediction results in each domain.
• Figure 3: Architectural detail of the Transformer block in the DenNet (denoising network).
• Figure 4: Process of extracting image patches and sinogram patches from the whole PET image.
• Figure 5: Visual comparison of SPET images produced using five different combinations of loss functions and network components shown in Table 1. The LPET image is shown on the left, followed by the results produced by the five different methods indicated in Table 1 (2nd-5th columns), and the ground truth (GT) in the last column. The corresponding difference maps between the produced results and the ground truth (GT) are depicted in the 2nd (axial view), 4th (sagittal view), and 6th (coronal view) rows. Red boxes show areas for detailed comparison.