Coordinated Reconstruction Dual-Branch Network for Low-Dose PET Reconstruction

Coordinated Reconstruction Dual-Branch Network for Low-Dose PET Reconstruction

---

연구 분야: Networking

---

---

학회: National Conference of Theoretical Computer Science

---

초록

Positron Emission Tomography (PET), known for its sensitivity and non-invasiveness in visualizing metabolic processes in the human body, has been widely utilized for clinical diagnosis. However, the procedure of PET imaging requires the administration of a radioactive tracer, which poses potential risks to human health. Reducing the usage of radioactive tracers leads to lower information content and increased independent noise. Therefore, the reconstruction of low-dose PET images becomes crucial. Existing reconstruction methods that learn a single mapping for low-dose PET reconstruction often suffer from over-denoising or incomplete information. To address this challenge, this work investigates the generation of realistic full-dose PET images. Firstly, we propose a simple yet reasonable low-dose PET model that treats each reconstructed voxel as a random variable. This model divides the reconstruction problem into two sub-problems: noise suppression and missing data recovery. Subsequently, we introduce a novel framework called the Coordinated Reconstruction Dual Branch Network (CRDB). The CRDB utilizes dual branches to separately perform denoising and information completion for PET reconstruction. Moreover, the CRDB leverages the Fast Channel Attention mechanism to capture diverse and unique information from different channels. Additionally, to emphasize pronounced distinctions, we adopt the Huber loss as the loss function. Quantitative experiments demonstrate that our strategy achieves favorable results in low-dose PET reconstruction.

---