Dual-branch residual encoder-decoder convolutional neural network (DB-REDCNN): a computed tomography-integrated multimodal network for positron emission tomography denoising.

Background: Positron emission tomography (PET) requires injection of radioactive tracers, which entails concerns regarding radiation exposure. Thus, there is a need for low-dose PET (LDPET) imaging that maintains full-dose PET (FDPET) quality while reducing tracer dosage. However, dose reduction often degrades image quality. Although deep learning methods have shown promise, denoising small lesions remains challenging. To address this, we propose a dual-branch network that incorporates structural information from computed tomography (CT) to enhance LDPET quality, preserve fine details, and improve small-lesion imaging. Methods: We propose the dual-branch residual encoder-decoder convolutional neural network (DB-REDCNN), a dual-branch multimodal network that leverages structural priors from paired CT images to enhance edge detail in PET reconstructions. Specifically, the network architecture consists of two parallel branches that independently extract modality-specific features from PET and CT inputs, followed by an effective fusion mechanism for reconstructing high-quality PET images. The overall architecture adopts a residual encoder-decoder design, with dedicated encoder and decoder modules at the input and output stages, respectively. It is important to note that although minor misalignments between CT and PET images may occur due to respiratory motion or patient movement, our dual-branch framework is inherently robust to such inconsistencies. For data acquisition, LDPET images were generated with an acquisition time of 30 seconds per bed position, while FDPET images were acquired over 150 seconds. For performance evaluation, we conducted extensive comparisons with several state-of-the-art deep learning methods. Moreover, the standardized uptake value (SUV) was included as an additional clinical indicator. Results: The proposed DB-REDCNN demonstrated superior performance across multiple evaluation metrics. In quantitative analysis, it achieved the largest improvements compared with LDPET images in root mean square error (0.0023±0.0075), peak signal-to-noise ratio (PSNR) (1.2314±4.8354), and the structural similarity index measure (0.0078±0.0153), with a significantly higher PSNR than REDCNN and consistent advantages across all three metrics (P<0.05). Edge sharpness assessments further confirmed the superiority of DB-REDCNN, which yielded the highest |K| values across five tumor slices with statistically significant improvements over competing models (P<0.01). In SUV evaluation, although the SUV_mean error of the DB-REDCNN was slightly higher than that of LDPET, the proposed method markedly reduced SUV_max error and outperformed all other approaches. Conclusions: The proposed DB-REDCNN model, by integrating structural priors from CT through a multibranch architecture, enhances the synthesis of FDPET from LDPET. It achieves superior quantitative performance, improves edge sharpness in lesion regions, and preserves peak SUV information, thereby offering a clinically valuable approach for radiation dose reduction in PET imaging.