Algorithm 1 Dual-path adversarial denoising network training procedure

Input:Clean image set $X=\{x_1,\dots ,x_n\}$. Noisy image set $Y=\{y_1,\dots ,y_n\}$ (unpaired). Hyper-parameters: $\alpha$, ${\tau }_D$, ${\tau }_G$, ${\lambda }_{\mathrm{gp}}$, ksize, batch_size, epochs, early_stop_patience  Parameters : ${\theta }_D$ — Denoiser (Dual-UNet) parameters. ${\theta }_G$ — Generator (UNet-G) parameters. ${\theta }_P$ — Discriminator (PatchGAN) parameters. Output: Trained denoiser $D(x^{\prime };{\theta }_D)$ with minimal noise and maximum detail preservation 1:Randomly initialize ${\theta }_D$, ${\theta }_G$, ${\theta }_P$ 2:Pre-compute Gaussian kernel $K=\mathrm{get\_gausskernel}(\mathrm{ksize},C=3)$ 3:for epoch $=1$ to epochs do 4:    for mini-batch $=1$to $\lceil |X|/\mathrm{batch\_size}\rceil$ do 5:        /* 1. Sample data */ 6:        $\left\{x\right\}\leftarrow$ Random batch from X 7:        $\left\{y\right\}\leftarrow$ Random batch from Y (same size) 8:        $\left\{z\right\}\leftarrow \mathcal{N}(0,1)$ noise map same spatial size as x 9:        /* 2. Forward passes */ 10:        $\widehat{x}\leftarrow y-D(y;{\theta }_D)$                                                                     ▹ Denoised image 11:        $\tilde{y}\leftarrow G(\mathrm{concat}(x,z);{\theta }_G)$                                                          ▹ Synthetic noisy image 12:        /* 3. Update Discriminator P */ 13:        ${\mathcal{L}}_P\leftarrow \alpha \left(P(\widehat{x},y;{\theta }_P)-P(x,y;{\theta }_P)\right)+(1-\alpha )\left(P(x,\tilde{y};{\theta }_P)-P(x,y;{\theta }_P)\right)+{\lambda }_{\mathrm{gp}}·\mathrm{gradient\_penalty}$ 14:        ${\theta }_P\leftarrow {\theta }_P-{\eta }_P{\nabla }_{{\theta }_P}{\mathcal{L}}_P$ 15:        if mini-batch mod num_critic $=0$ then 16:           /* 4. Update Denoiser D */ 17:           ${\mathcal{L}}_D\leftarrow -(1-\alpha )P(\widehat{x},y;{\theta }_P)+{\tau }_D{\parallel \widehat{x}-x\parallel }_1$ 18:           ${\theta }_D\leftarrow {\theta }_D-{\eta }_D{\nabla }_{{\theta }_D}{\mathcal{L}}_D$ 19:           /* 5. Update Generator G */ 20:           ${\mathcal{L}}_G\leftarrow -\alpha P(x,\tilde{y};{\theta }_P)+{\tau }_G{\parallel \mathrm{GF}(\tilde{y}-x)-\mathrm{GF}(y-x)\parallel }_1$ 21:           ${\theta }_G\leftarrow {\theta }_G-{\eta }_G{\nabla }_{{\theta }_G}{\mathcal{L}}_G$ 22:        end if 23:    end for 24:    /* 6. Early stopping check */ 25:    val_loss ← MAE on validation set 26:    if val_loss has not improved for early_stop_patience epochs then 27:        break 28:    end if 29:end for 30:return  $D(x^{\prime };{\theta }_D)$