Research on noise reduction technology of PCD spectral imaging based on Transformer

Spectral computed tomography (CT) utilizing photon counting detector (PCD) has advanced quickly in recent years, and its utilization in clinics has been gradually increasing. Compared with traditional CT, PCD-CT has the advantages of multi substances quantification, and can significantly reduce image noise and scanning dose while maintaining image quality. However, due to the limitation of hardware, PCD is affected by photon starving, charge sharing effect and pulse superposition effect, leading to statistical fluctuation noise in spectral CT reconstruction. Nowadays, deep learning has become the state-of-the-art method in noise suppressing of PCD-CT. In this work, we introduced the attention mechanism based on Transformer into CNN backbone and proposed a new reconstruction U-Net. This architecture can effectively explore both global and local features of noisy maps. The qualitative and quantitative experiments indicated that our method can effectively improve the performance in noise reduction and spectral CT reconstruction.