Why do State-of-the-art Super-Resolution Methods not work well for Bone Microstructure CT Imaging?

3D Computerized Tomography (CT) is a gold stan-dard technique to assess bone microstructure in the context of bone diseases such as osteoporosis. However, when acquired in-vivo, bone images may suffer from a low spatial resolution and the presence of noise due to the limited tolerable radiation exposure. One way to overcome this issue consists in applying Super-Resolution (SR) techniques that aim at recovering high resolution images. Significant progress has been recently made thanks to deep learning SR methods trained on natural image datasets. To measure the reconstruction quality, Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM) are commonly used in the SR literature. In this paper, we give evidence of the limitation of these two criteria. Through extensive experiments performed from a dataset of mice tibias specifically collected and imaged for this study, we show that state of the art deep learning-based SR methods miss important details about the bone microstructure which is not reflected by the PSNR and SSIM values. This study opens the door to future promising lines of research including new SR methods regularized with respect to morphometric and tonological parameters of bone microstructures.