[论文解读] Uncertainty-guided Generation of Dark-field Radiographs

Lina Felsner, Henriette Bast+5 more|arXiv (Cornell University)|Jan 22, 2026

Advanced X-ray Imaging Techniques被引用 0

一句话总结

论文提出一个渐进式 GAN 框架，在对标准衰减胸部放射影像生成 X 射线暗场图像的同时，对预期内在不确定性（aleatoric）与知识不确定性（epistemic）进行建模，以提升可靠性和可解释性。

ABSTRACT

X-ray dark-field radiography provides complementary diagnostic information to conventional attenuation imaging by visualizing microstructural tissue changes through small-angle scattering. However, the limited availability of such data poses challenges for developing robust deep learning models. In this work, we present the first framework for generating dark-field images directly from standard attenuation chest X-rays using an Uncertainty-Guided Progressive Generative Adversarial Network. The model incorporates both aleatoric and epistemic uncertainty to improve interpretability and reliability. Experiments demonstrate high structural fidelity of the generated images, with consistent improvement of quantitative metrics across stages. Furthermore, out-of-distribution evaluation confirms that the proposed model generalizes well. Our results indicate that uncertainty-guided generative modeling enables realistic dark-field image synthesis and provides a reliable foundation for future clinical applications.

研究动机与目标

• Motivate generation of dark-field images from conventional 2D X-rays to enable large-scale data augmentation.
• Introduce an uncertainty-guided progressive GAN framework that jointly models aleatoric and epistemic uncertainty.
• Evaluate robustness and generalization of the approach on in-house dark-field data and out-of-domain datasets.

提出的方法

• Use a three-stage progressive GAN with uncertainty-guided refinement where aleatoric estimates act as attention maps for subsequent stages.
• Model aleatoric uncertainty per pixel with parameters alpha (scale) and beta (shape) of a generalized Gaussian distribution.
• Estimate epistemic uncertainty via Monte Carlo dropout with multiple stochastic forward passes at inference.
• Incorporate a residual consistency loss to align noise textures with realistic dark-field statistics.
• Train with data augmentation and a cosine-annealing learning rate schedule on paired attenuation-dark-field images.
• Evaluate using MSE, SSIM, and PSNR across stages to assess fidelity.

Figure 2: Proposed Uncertainty-Guided Progressive GAN framework, illustrated for model training at stage three. At this stage, the previous layers are frozen. The aleatoric uncertainty estimates are used as attention maps to guide the model’s weights for refinement in the subsequent stage. Dropout i

实验结果

研究问题

• RQ1 Can a GAN translate attenuation chest radiographs into realistic dark-field images?
• RQ2 Does explicitly modeling aleatoric and epistemic uncertainty improve quality, reliability, and interpretability of generated dark-field images?
• RQ3 How does progressive refinement affect quantitative fidelity and uncertainty maps across stages?
• RQ4 Does the model generalize to out-of-distribution chest X-rays beyond the grating-based acquisition domain?

主要发现

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

• Generated dark-field images show high visual and quantitative alignment with real dark-field data across stages.
• MSE decreases, while PSNR and SSIM increase progressively: Stage 1 (MSE 0.0131, PSNR 19.35, SSIM 0.38); Stage 2 (0.0125, 19.57, 0.47); Stage 3 (0.0123, 19.71, 0.52).
• Aleatoric uncertainty generally decreases with stage, with higher values concentrated in lung regions; epistemic uncertainty remains relatively constant across stages.
• Out-of-distribution NIH Chest X-ray data yield realistic dark-field generation with corresponding uncertainty indicating unfamiliar regions, including a notable failure case with artifacts and higher uncertainty.

Figure 3: Comparison of attenuation, real dark-field, and generated dark-field images for two patients.