Fig. 4: 3D ex vivo virtual staining results of a skin tissue area around the dermal-epidermal junction and their comparison with ground truth, actual acetic acid staining.
From: Biopsy-free in vivo virtual histology of skin using deep learning
a–d Label-free RCM images showing an ex vivo skin tissue area at different depths around dermal-epidermal junction without any staining, served as the network inputs. The depth of (b), (c), and (d) were 12.16, 24.32, and 36.48 μm below a into the skin, respectively. e Cross-section of the RCM image stack of the tissue area including (a–d). Lines in different colors are used to indicate the depth positions of (a–d). f–i Acetic acid virtual staining results of the same tissue area and depth as (a–d) generated by the deep neural network VSAA. j is the image stack cross-section of the acetic acid virtual staining results including (f–i) generated using the acetic acid virtually stained tissue images. k–n Pseudo-H&E virtual staining results generated using the acetic acid virtually stained tissue images (f–i). These H&E-like images were generated by the pseudo-H&E virtual staining network VSHE that took both the RCM images of the unstained tissue (a–d) and acetic acid virtually stained tissue images (f–i) as input (see solid arrows below the upper panel). o Cross-section of the pseudo-H&E virtually stained tissue image stack including (k–n). u–x RCM images of the same tissue area and depth as (a–d) after the actual acetic acid staining process, served as ground truth for (f–i). y shows the cross-section of the image stack of the tissue stained with acetic acid including (u–x). p–s Pseudo-H&E virtual staining results generated using the actual acetic acid-stained images (u–x). These H&E-like images were generated by the same pseudo-H&E virtual staining network VSHE that took the RCM images of the unstained tissue (a–c) and actual acetic acid-stained images (q–s) as input (see dashed arrows below the upper panel and see Materials and Methods for more details). t shows the cross-section of the pseudo-H&E virtually stained tissue image stack including (p–s) generated using the actual acetic acid-stained images. Zoomed-in views of some portions of the images are provided at the bottom for a better visual comparison of details.
