Bayesian Joint Reconstruction and Uncertainty Quantification for Limited-Data Multi-Contrast X-Ray CT

Abstract

Multi-contrast X-ray computed tomography (CT) combines attenuation contrast with phase and dark-field contrast, which are sensitive to refraction and ultra-small-angle scattering, respectively. Common implementations record several intensity measurements per projection angle by phase or mask stepping, which lengthens acquisition and, at a fixed exposure per measurement, raises the dose. Limited-data acquisition via sparse angular sampling or reduced stepping lowers these costs but makes the data less informative, calling for prior information and uncertainty quantification. We study joint Bayesian reconstruction of attenuation, phase, and dark-field images from separately retrieved sinograms. A joint total variation term in the prior couples edge locations across the images but not their values. Posterior sampling with the No-U-Turn Sampler complements maximum a posteriori estimates with pixel-wise uncertainty. We compare this joint prior with channel-wise total variation alone on synthetic and experimental Talbot-Lau interferometry data under full, sparse-angle, sparse-step, and combined acquisition (up to 24 times fewer exposures). Both maximum a posteriori reconstructions suppressed undersampling artifacts, with synthetic relative errors 1.6-11 times lower than those of filtered backprojection. The joint prior gave modest synthetic accuracy gains, a cleaner experimental dark-field background, 23-30 % lower mean dark-field posterior standard deviation, and 4-38% higher median effective sample size in experimental limited-data runs. In the synthetic study, posterior standard deviations matched or exceeded the errors in flat regions but underestimated them at boundaries, especially in phase. Experimental phase chains mixed slowly under angular undersampling. The formulation also applies to edge illumination with adapted forward and noise models.