Fig. 2.

The overall structure of our unified motion correction and denoising network (MDPET). The reference gate low-dose image L_ref and N-th gate low-dose images L_n are fed into each Siamese Pyramid Network (SP-Net) within our Temporal Siamese Pyramid Network (TSP-Net). The predicted transformation fields T_n simultaneously transform the paired L_n and H_n. The transformed low-dose gated image ${\widehat{L}}_n$ are averaged and subsequently fed into the denoising network for denoising. Our MDPET is trained in a unified fashion with registration loss ${\mathcal{L}}_{\mathit{\text{reg}}}$, denoising loss ${\mathcal{L}}_{DN}$, and adversarial loss ${\mathcal{L}}_{\mathit{\text{adv}}}$ combined.