Algorithm 1. Self-correction learning mechanism

Input:${\widehat{\omega }}_0$: initialized model weight by $f_{\mathit{enc}}$, $g_{\mathit{md}}$, $g_{\mathit{ad}1}$, $g_{\mathit{ad}2}$, and $\phi$ after domain adaptation; ${\widehat{y}}_0^a$: augmented initial pseudo label generated by ${\widehat{\omega }}_0$; $\mathit{Ep}$: epoch in each cycle; $C$: total number of cycles;

Output: Optimal ${\widehat{\omega }}_C$

1: Construct a training dataset ${\mathcal{D}}_{\mathit{training}}$ using ${\mathcal{D}}_S=\left(I_S,Y_S\right)$ and ${\mathcal{D}}_T=\left(I_T,{\widehat{Y}}_0^a\right)$

2: for each $c\in [1,C]$do

3:  for each $e\in [1,\mathit{Ep}]$do

4:   Update the learning rate;

5:   Calculate loss $L_{\mathit{seg}}$ by Eq. (7);

6:  end for

7:  Update weight ${\widehat{\omega }}_{c+1}$ for initializing the weight of the next cycle $c+1$ by Eq. (9);

8:  Generate pseudo label ${\widehat{Y}}_{c+1}$ using ${\widehat{\omega }}_c$;

9:  Augment pseudo label ${\widehat{Y}}_{c+1}^a$ using ${\widehat{\omega }}_c$;

10:  Pseudo label refinement by Eq. (10);

11:  Re-construct a training dataset ${\mathcal{D}}_{\mathit{training}}$ using ${\mathcal{D}}_S=\left(I_S,Y_S\right)$ and ${\mathcal{D}}_T=\left(I_T,{\widehat{Y}}_{c+1}^a\right)$

12: end for