3D-реконструкция кожи и пространственное картирование плотности иммунных клеток, расстояния между сосудами и эффектов воздействия солнца и старения.

For 3D reconstruction of the serial sections, we first manually select one reference 2D AF image from the stack of 26 serial images (Fig. 3a, b and Methods). All AF images used for reconstruction are unstained (i.e., the image is taken prior to any marker staining), hence AF-based registration is independent of biomarker signal or signal variations associated with staining. Compared to other registration approaches that have been used for 3D reconstruction14, we automatically segment AF in each serial image using Otsu thresholding, morphological closing, and retaining the largest component39 (Fig. 3c) to generate a 2D AF mask. The 2D AF mask focuses registration on a region of interest (ROI) and filters out background noise or artefacts that may interfere with the registration process. Post serial image registration, 3D volumes of AF and cells are created, and 3D connectivity of all cells is used to fuse overlapping cells in adjacent serial sections to prevent overcounting (Fig. 3d, e). We use ITK’s 3D connected component image filter to merge overlapping cells and classify cells in 3DClass Template Reference. https://itk.org/Doxygen/html/classitk_1_1ConnectedComponentImageFilter.html (2023)." href="/articles/s42003-023-04991-z#ref-CR40" id="ref-link-section-d87231388e979"40,41. Connected component image filter has historically been used in merging segmentations in 3D42,43,44 and for refining cell segmentation in 2D45. To the best of our knowledge, this is the first time a 3D connected component has been used to fuse overlapping cells in 3D in serial histological sections. Mean dice similarity coefficient (DSC)46 was used to assess image registration accuracy, and we found an overlap accuracy of 0.95 ± 0.04 for 24 serial sections in the ten reconstructed skin volumes (the last two sections were used for deep learning training and excluded). An overlap DSC of 0.90 is classified as high quality in the 3D registration or reconstruction community14. The normalized cross correlation between the serial AF sections (a metric of registration quality) was 0.6 ± 0.07 for all AF serial sections. This result is comparable to established 3D reconstruction methods13 and is good considering the deformation/damage that can occur during cyclic multiplexing, which could negatively affect accurate registration. Slides were also randomly picked from each of the reconstructed skin volumes and segmented cells and biomarkers were overlaid on the images for visual validation. To mitigate potential issues associated with higher or lower signal intensity for a slide, any discrepancies in cell density of a biomarker were identified and a higher or lower probability threshold was used on the probabilistic segmentation to improve segmentation. Such manual biomarker probability adjustments were necessary in less than 2% of the whole slide images dataset./p>

For 3D reconstruction a reference AF image was manually identified from 24 serial sections of every region based on image quality (contrast, wear and tear, deformation etc.). All 2D AF images were masked (with a tissue mask) and all 2D AF serial section images were registered to the chosen reference image. Otsu thresholding uses intra-class variance computed from image histogram to set the thresholds for the AF background and the foreground. Hence while the AF signal distribution may change from one serial section to another, the threshold is set based on intra-class variance and is automatically adjusted from one serial section to another to maximize the separation of the background and the foreground. Our registration model uses local patch-based, normalized cross correlation to establish correspondences making our model fast without compromising on accuracy. Patch wise normalized cross correlation depends less on absolute intensity values and depends more on a good separation of the background and foreground signal which we obtain by masking our autofluorescence image. After affine registration, B-spline based deformable registration68, we use normalized mutual information to mitigate image intensity differences that may occur between one serial section to another in the autofluorescence images. A block matching strategy69 was adopted to determine the transformation parameters for the affine registration from masked AF images (Fig. 3c). The similarity between a block from the reference AF was computed relative to the AF serial sections. The best corresponding block defined the displacement vector for the affine transformation70. Normalized mutual information was chosen as the similarity matrix and maximized to achieve the registration. The transformation map obtained from the registration of the AF images was applied to individual biomarkers for all serial sections to create a 3D volume of endothelial, T killer, T reg, T helper cells and macrophages (Fig. 3d) and overlaid on 3D AF volume as shown in Fig. 3e. Post registration, 3D connectivity of all cells were used to fuse overlapping cells in adjacent serial sections to prevent overcounting. Cells overlapping in 3D in adjacent sections are automatically connected and considered as a single entity in 3D using ITK’s 3D connected filterClass Template Reference. https://itk.org/Doxygen/html/classitk_1_1ConnectedComponentImageFilter.html (2023)." href="/articles/s42003-023-04991-z#ref-CR40" id="ref-link-section-d87231388e1796"40. ITK is an opensource software widely used for both 2D and 3D medical image analysis. ITK’s 3D connected filter is used to merge binary labels in 3D based on overlap of the labels in 3D (adjacent sections post 3D reconstruction). While the Watershed algorithm is often used for merging labels in 3D, it requires multiple parameters (diffusion, gradient, thresholding) to be tuned manually for a dataset that may not generalize in a large dataset like ours. This would result in merging of un-related cells (due to diffusion parameter) or cell dropping (due to gradient threshold parameter). Instead, a connected component filter has been used to merge segmented cells that are locally connected in 2D to create a single unit for each cell. Here, we extend that framework to 3D. No manual tuning of the parameters is required for this connected component filter approach, and ITK’s 3D connected filters are routinely used in medical image analysis. All code can be found at GitHub - hubmapconsortium/MATRICS-A: Multiplexed Image Three-D Reconstruction and Integrated Cell Spatial -Analysis, the docker container/environment to run the code can obtained by using docker pull hubmap/gehc:skin and test data for skin region 7 can be found at Human Digital Twin: Automated Cell Type Distance Computation and 3D Atlas Construction in Multiplexed Skin Biopsies | Zenodo. There is a corresponding ReadMe file that provides context and instructions for the repository’s contents and could be found here MATRICS-A/README.md at main · hubmapconsortium/MATRICS-A · GitHub./p>

Class Template Reference. https://itk.org/Doxygen/html/classitk_1_1ConnectedComponentImageFilter.html (2023)./p>