Uncertainty-Guided Progressive GAN

• The paper introduces uncertainty-guided progressive GANs that use a coarse-to-fine synthesis strategy with stagewise refinement and uncertainty attention to improve image fidelity.
• It combines adversarial, reconstruction, and residual consistency losses to yield robust image translation while modeling both aleatoric and epistemic uncertainties.
• Quantitative results demonstrate improved PSNR/SSIM and reduced MSE across stages, highlighting the framework’s effectiveness for data-limited clinical imaging tasks.

Uncertainty-Guided Progressive Generative Adversarial Network (UG-ProgGAN and UP-GAN) is a specialized framework in medical image synthesis and translation, integrating generative adversarial learning with explicit modeling of both aleatoric and epistemic uncertainty within a progressive growing paradigm. The design enables high-fidelity image generation, robust uncertainty quantification, and targeted refinement, and is particularly suited to data-limited clinical imaging tasks such as dark-field radiograph synthesis and multimodal translation (Felsner et al., 22 Jan 2026, Upadhyay et al., 2021).

1. Architectural Principles

Uncertainty-Guided Progressive GANs employ a multi-stage architecture in which generator–discriminator pairs operate at sequentially increasing spatial resolutions. Each stage $s$ consists of a generator $G_s$ and a discriminator $D_s$:

• Stagewise progression: At each stage ($s=1,2,3$), the generator and discriminator are trained at resolutions, e.g., $64\times64$, $128\times128$, and $256\times256$. After a stage completes, its parameters are frozen; subsequent stages add “refinement” layers to focus on residual structure and fine details, thereby implementing a coarse-to-fine synthesis scheme (Felsner et al., 22 Jan 2026).
• Generator design:
  • Stage 1: Input is the source image (e.g., attenuation X-ray); the network outputs a preliminary synthesis plus pixelwise aleatoric parameters $(\alpha, \beta)$ and uses dropout for epistemic estimation.
  • Stages $s>1$: The generator receives as input the source, the previous stage’s output, and an uncertainty map as an attention channel, focusing refinement on high-uncertainty regions.
• Discriminator: PatchGAN-style discriminators $D_s$ take a concatenation of source and target (real or fake) and produce a map of patch-level real/fake logits, as in Isola et al. (2017).

2. Losses and Optimization Objectives

The total loss for each generator stage is a weighted sum:

$$\mathcal{L}^{(s)}_{\mathrm{total}} = \mathcal{L}^{(s)}_{\mathrm{GAN}} + \lambda_{\mathrm{rec}}\,\mathcal{L}^{(s)}_{\mathrm{rec}} + \lambda_{\mathrm{res}}\,\mathcal{L}^{(s)}_{\mathrm{res}}$$

• Adversarial loss (Least Squares GAN):

  $$\mathcal{L}^{(s)}_{\mathrm{GAN}} = \tfrac12\, {\mathbb E}_{x,y}\bigl[(D_s(x,y)-1)^2\bigr] + \tfrac12\, {\mathbb E}_{x}\bigl[(D_s(x,\hat y_s)+1)^2\bigr]$$

• Reconstruction loss (typically $L_1$):

  $$\mathcal{L}^{(s)}_{\mathrm{rec}} = {\mathbb E}_{x,y}\bigl[\|y - \hat y_s\|_{1}\bigr]$$

• Residual consistency loss (texture-regularizing):

  $$R(u) = u - \mathrm{Blur}(u), \quad \mathcal{L}^{(s)}_{\mathrm{res}} = {\mathbb E}_{x,y}\bigl[\|R(y) - R(\hat y_s)\|_1\bigr]$$

• Uncertainty-guided negative log-likelihood term (in UP-GAN (Upadhyay et al., 2021)):

  $$\mathcal{L}_{\text{uncertainty}} = -\frac{1}{K}\sum_{i,j} \log p(\epsilon_{ij}|\alpha_{ij},\beta_{ij})$$

  where the density is that of the Generalized Gaussian (see Section 3).

Hyperparameters for $\lambda_{\mathrm{rec}}, \lambda_{\mathrm{res}}$, and, where applicable, $\lambda_{\mathrm{unc}}$ are set as reported in the respective studies.

3. Uncertainty Modeling: Aleatoric and Epistemic

The uncertainty-guided approach models and exploits two types of uncertainty:

• Aleatoric uncertainty: For each pixel $(i, j)$, the generator infers scale $\alpha_{ij}$ and shape $\beta_{ij}$ of a generalized Gaussian, modeling observation noise or ambiguity. The aleatoric (data) uncertainty at pixel $(i, j)$ is

  $$\sigma_{ij} = \alpha_{ij} \sqrt{ \frac{\Gamma(3/\beta_{ij})}{\Gamma(1/\beta_{ij})} }$$

  These $\sigma_{ij}$ maps are input as attention weights in later stages, directing the refinement network to focus on structurally uncertain or ambiguous regions (Felsner et al., 22 Jan 2026, Upadhyay et al., 2021).

• Epistemic uncertainty: Modeled via Monte Carlo dropout, where dropout is active at inference and $T$ stochastically sampled outputs $\{\hat y_s^t\}_{t=1}^T$ are generated. Epistemic uncertainty at pixel $(i, j)$ is estimated by sample variance:

  $\mathrm{Var}_t[\hat y_s^t(i,j)]$

  This enables the identification of model uncertainty arising from limited data or distributional shift.

In UP-GAN (Upadhyay et al., 2021), only aleatoric uncertainty is modeled; epistemic components are highlighted as a future direction.

4. Training Protocols and Implementation

• Data: For dark-field synthesis, 269 paired attenuation/dark-field chest radiographs (split 227/15/27 for train/val/test); for multimodal translation (e.g., PET→CT, undersampled MRI), datasets as described in (Upadhyay et al., 2021).
• Augmentation: Spatial transforms and intensity jittering for robustness.
• Optimization: Adam optimizer; learning rates and cosine annealing schedules per stage ($8 \times 10^{-6}$ for dark-field, $2 \times 10^{-3}$ for UP-GAN) with batch sizes set by hardware capacity.
• Progressive scheme: Networks are trained sequentially per stage, with previously learned layers frozen, then optionally fine-tuned jointly (UP-GAN).
• Dropout: Rate 0.1 in generators for uncertainty estimation; at test time, MC sampling ($T=20$) is performed for epistemic evaluation.

5. Evaluation and Quantitative Results

Evaluation uses structural and fidelity metrics:

Stage	MSE	PSNR (dB)	SSIM
1	0.0131±0.0067	19.35±2.14	0.38±0.06
2	0.0125±0.0066	19.57±2.24	0.47±0.05
3	0.0123±0.0067	19.71±2.37	0.52±0.05

Metrics improve monotonically with each progressive stage, confirming the advantage of coarse-to-fine refinement (Felsner et al., 22 Jan 2026). Qualitative results show high visual fidelity between real and synthesized images; uncertainty maps highlight areas of model uncertainty. Out-of-distribution testing demonstrates robustness, with uncertainty spikes at anatomical or device configurations unseen in training.

In UP-GAN (Upadhyay et al., 2021), full and weak-supervision settings for PET→CT, MRI reconstruction, and MRI motion correction similarly demonstrate that uncertainty guidance increases PSNR/SSIM and robustness, outperforming baselines (pix2pix, PAN, MedGAN). Removal of uncertainty attention yields significant drops in performance.

6. Significance of Progressive and Uncertainty Guidance

• Progressive Growing: Drives substantial gains in structural image quality (PSNR/SSIM). Early stages capture coarse structure, later stages refine anatomy and texture.
• Uncertainty Attention: Aleatoric maps enable spatially adaptive focus on difficult or noisy regions (e.g., lung periphery, motion artifacts), resulting in sharper reconstructions and reduced error. A reported $\sim6\%$ reduction in MSE from stage 1 to 3 is attributed to this mechanism (Felsner et al., 22 Jan 2026).
• Epistemic Uncertainty Utility: Facilitates detection of out-of-distribution samples and model failure, as uncertainty estimates are elevated in problematic regions. This provides an unsupervised reliability signal.

A plausible implication is that such frameworks offer not only higher image fidelity but also essential uncertainty measures for clinical decision support and expert triage.

7. Applications, Limitations, and Future Directions

Applications include:

• Synthetic dark-field radiograph generation from standard X-rays (Felsner et al., 22 Jan 2026).
• Multi-modal image-to-image translation (PET→CT, MRI, etc.) (Upadhyay et al., 2021).

Limitations noted are:

• Current implementations model primarily aleatoric uncertainty; joint aleatoric–epistemic modeling, ensemble approaches, and deployment to 3D or temporal imaging are potential future directions (Upadhyay et al., 2021).
• While high performance is demonstrated on relatively small datasets, scaling to large, multi-site cohorts and integration into clinical pipelines requires further validation.

The evidence supports the role of Uncertainty-Guided Progressive GANs as an effective and robust approach to medical image synthesis, offering reliable uncertainty estimates alongside improved image quality.