Algorithm 1.

1: Input: A set of training low-dose PET images ${\mathbf{I}}^L=\left\{I_1^L,I_2^L,\dots ,I_N^L\right\}$, a set of training multimodal MRI images including ${\mathbf{I}}^{T1}=\left\{I_1^{T1},I_2^{T1},\dots ,I_N^{T1}\right\}$, ${\mathbf{I}}^{FA}=\left\{I_1^{FA},I_2^{FA},\dots ,I_N^{FA}\right\}$, ${\mathbf{I}}^{MD}=\left\{I_1^{MD},I_2^{MD},\dots ,I_N^{MD}\right\}$, and a set of training full-dose PET images ${\mathbf{I}}^F=\left\{I_1^F,I_2^F,\dots ,I_N^F\right\}$. N is the total number of training samples.

2. Perform the 3D LA-GANs between IL, IT1, IFA, IMD and IS to obtain the fusion network F, the generator network G, and the discriminator network D.

3. For each training sample i (i=1,2,...,N), use the above trained fusion network F and generator network G to generate the synthetic full-dose PET image ${\tilde{I}}_i^F$. Finally, get the estimations for all training subjects ${\tilde{\mathbf{I}}}^F$.

4. The synthesis of the training subjects ${\tilde{\mathbf{I}}}^F$, along with the original low-dose PET IL and multimodal MRI images IT1, IFA and IMD, are all input to the subsequent LA-GANs network, namely, auto-context LA-GANs, to obtain the updated fusion network F′, the generator network G′ and the discriminator network D′.

5: Output: The trained fusion networks for LA-GANs and auto-context LA-GANs F and F’, as well as the trained generator networks G and G’.