Unsupervised Super Resolution in X-ray Microscopy using a Cycle-consistent Generative Model

Abstract

X-ray microscopy (XRM) is a tomographic imaging modality that has gained interest in the context of understanding bone-related diseases on the micro scale due to its high spatial resolution and strong bone to soft tissue contrast. Although in-vivo imaging of bone structures on the micro scale is desired from a medical perspective, high radiation dose so-far prohibits imaging living animals. Research has been focused on generating high-quality reconstructions while maintaining a low X-ray dosage. However, low dose acquisitions result in noisy images with a lower resolution. This study focuses on using an unsupervised deep-learning approach to accurately reconstruct high-resolution (HR) XRM images from their noisy low-resolution (LR) counterparts. We consider an unsupervised approach in a general case where paired data (low-/high resolution pairs) are unavailable. We propose the use of a cycle-consistent generative adversarial network (GAN) for this super resolution task which is to learn the mapping from noisy LR to HR images. Quantitative and qualitative assessments show that our method produces accurate high-resolution XRM reconstructions from their noisy low-resolution counterparts, increasing the peak signal-to-noise ratio (PSNR)/structural similarity index (SSIM) from 18.15/0.52 (baseline) to 31.94/0.73 (proposed method). We believe that our proposed XRM super resolution pipeline provides a valuable tool toward high-resolution in-vivo XRM imaging.