Air artifact suppression in phase contrast micro-CT using conditional generative adversarial networks

3D virtual histology of formalin-fixed and paraffin-embedded (FFPE) tissue by means of phase contrast micro-computed tomography (micro-CT) is an increasingly popular technique, as it allows the 3D architecture of the tissue to be addressed without the need of additional heavy ion based staining approaches. Therefore, it can be applied on archived standard FFPE tissue blocks. However, one of the major concerns of using phase contrast micro-CT in combination with FFPE tissue blocks is the trapped air within the tissue. While air inclusion within the FFPE tissue block does not strongly impact the workflow and quality of classical histology, it creates serious obstacles in 3D visualization of detailed morphology. In particular, the 3D analysis of structural features is challenging, due to a strong edge effect caused by the phase shift at the air-tissue/paraffin interface. Despite certain improvements in sample preparation to eliminate air inclusion, such as the use of negative pressure, it is not always possible to remove all trapped air, for example in soft tissues such as lung. Here, we present a novel workflow based on conditional generative adversarial networks (cGANs) to effectively replace these air artifact regions with generated tissue, which are influenced by the surrounding content. Our results show that this approach not only improves the visualization of the lung tissue but also eases the use of structural analysis on the air artifact-suppressed phase contrast micro-CT scans. In addition, we demonstrate the transferability of the generative model to FFPE specimens of porcine lung tissue.