Visualization of pathologies, segmentation, and quantification
The overview scans of all paraffin embedded Covid-19 positive samples and one biopsy of a hydrated control lung were analyzed in terms of structural characteristics. The top row of Figure 6 shows the workflow of the analysis on the example of the 3D reconstruction of sample V. Based on the tissue mask the distances for each tissue voxel to the next voxel containing oxygen were determined. For the analysis of the tissue it is mandatory to consider the three-dimensionality of the samples, similar to the analysis of porous structures (Müller et al., 2002). More generally, we claim that 3D histology analysis based on x-ray tomography and light sheet microscopy (Power and Huisken, 2017) is required to quantitatively understand functional tissue properties based on its 3D architecture. Here, this can also be seen in Figure 6d, which shows a zoom of the distance analysis around a small blood vessel which is marked with a yellow box. While the wall thickness appears quite homogeneous in the 2d slice, the distance analysis reveals that the vessel is thicker on the top right.
Figure 6.  Tissue compactness and distance metrics: The 3D reconstructions of the lung tissue were analyzed in terms of specific surface and characteristic lengths.
The workflow is sketched for sample V in the top part (a–g). In a first step, the tissue was segmented using Ilastik. A slice of the reconstructed electron density is shown in (a). Based on the segmentation, the areas of air which are directly connected and potentially filled with oxygen (or blood) are masked out (b). (c) For each of the remaining tissue voxels the shortest possible distance to air was calculated. Especially around vessels and larger alveoli, the distances are larger. (d) Zoom into an area around a vessel. Further analysis is based on the distance distribution shown in (e). (f) Volume rendering of the reconstructed electron density, with (g) showing the corresponding 3D distance map. (h) Based on the tissue segmentation of all samples, the distance from the tissue interior to the closest air compartment was calculated. In order to compare all samples, the count of voxels was normalized by the total volume of the respective sample. The specific surface area SV (represented by the first value of each curve), the characteristic length Lc and the mean distance dO2¯ for each sample was calculated based on this data. Double logarithmic scale, bin width of the distribution of distances: 1 v⁢x. Scale bars: (a–c) 100⁢μ⁢m. The distribution of the distances obtained for a given slice is shown in Figure 6e. The swelling of the alveolar walls as well as the inflamed blood vessels can be identified by comparing the reconstructed electron density and the 3D-distance map (see Figure 6f and g). Figure 6h shows the distribution of the tissue-air distances (histogram) for all samples, following the workflow illustrated in Figure 6e. The binning of distances was set to one voxel length. The figure underlines the high diversity of the tissue structure which could already be seen in the 3D histology. Further, it directly informs about the specific surface SV, which is given by the first point of the graph. The corresponding parameters and metrics are tabulated in Table 1 for all samples. Additionally, the mean concentration of lymphocytes cl within the lung tissue is listed for all samples. The values quantify the general structure of the tissue which is qualitatively discernible by eye. Samples with a high amount of swollen, inflamed blood vessels and thick hyaline membranes exhibit a larger characteristic length. Note, that the control lung was prepared in a hydrated environment and shrinking due to further preparation of the sample does not occur. Hence, the results cannot be directly compared to the paraffin embedded samples. Further, the analysis of the lymphocyte concentration was performed since no lymphocytes were found in the reconstructed volume. The low values of Lc and dO2¯ for sample II correlate with the lack of ground-glass opacification in clinical CT. Based on the extracted structural parameters, the degree of inflammation and swelling of lung tissue can be evaluated. E.g. patient II has the highest surface area volume-ratios while sample I and VI have a relatively low specific surface. Larger characteristic lengths may also be indicative of inflammation and the formation of hyaline membranes, which will be evaluated in the following based on ROI and high-resolution reconstructions.
Table 1.  Results of the analysis of tissue characteristics: specific surface area SV, characteristic length Lc and mean distance dO2¯ and standard deviation from all tissue voxels to air as well as the mean concentration of lymphocytes cl for all six Covid-19 positive samples as well as for one control sample. Colors match the distance graphs in Figure 6.
Patient no.  SV (%)  Lc (μm)  dO2¯ (μm)  c l     ( 10 5  /  m m  3  )
I  13.87  9.4  5.9±5.3  16.0
II  46.56  2.8  2.1±1.0  14.1
III  33.75  3.9  2.5±1.5  4.4
IV  25.90  5.0  3.2±2.3  7.1
V  19.28  6.7  3.6±2.1  4.8
VI  11.87  11.0  9.1±10  6.1
CTRL (hyd.)  20.04  6.5  5.0±5.1  - Figure 7 illustrates the aggregation of hyaline membrane in the vicinity of a single alveole. Volumetric renderings in Figure 7a and b demonstrate particular attachment of fibrin to the alveolar walls. In cases of severe hyaline membrane formation as for this patient, this pathological alteration can be tracked throughout the volume Figure 7c–e. In Figure 7f, hyaline membranes of neighboring alveoles are indicated. In the 3D-context, their locations with respect to blood vessels can be inspected, see Figure 7g, which exemplifies a direct connection of hyaline membranes to the vasculature.
Figure 7.  Rendering of hyaline membrane attached to alveolar walls (patient V, parallel beam-scan of a 1 mm punch).
The rendered subvolume was restricted to 1.15×1.10×0.56 mm3, to contain a single alveole foremost. (a) Volume rendering of the segmented hyaline membrane in same spatial orientation as (c)-(e), which show virtual slices through the (c) top, (d) center and (e) bottom of the alveole. For a better spatial classification, (b) gives a combination of the volume in (a) and the slice in (d). (f) Volume rendering of the entire subvolume including neighboring alveolae. (g) Zoom-in onto a major blood vessel (red) which is directly connected with the hyaline membrane. Scale bars: (c–e) 300⁢μ⁢m. The severeness of hyaline membrane formation is case-specific, as the yellow rendering in Figure 8a,b,d demonstrates reduced amounts of deposits for patient I in a subvolume of parallel-beam reconstructions. Further, lymphocytes (red) were identified based on the automated cell segmentation (see Materials and methods). For clearer visualization, each cell is rendered as a sphere with a size corresponding to the mean cell volume in Figure 8a,b,d. Based on convolution of the cell positions with a sphere of 100⁢μ⁢m in radius, the local cell density was calculated (Töpperwien et al., 2018) and presented as 3D-maps of cells/mm3 in Figure 8c,e. This concept was then translated to a 3D-stitched volume of an entire tissue block as shown in Figure 8f,g.
Figure 8.  Rendering of alveolar wall with hyaline membrane and quantification of lymphocyte infiltration (unstained tissue, patient I, parallel beam scan of a 1 mm punch).
The illustrations in (a–e) show a subvolume of 0.60×0.48×0.26 mm3, (f and g) this concept applied to the full tissue block of 2.57×2.99×0.98 mm3. (a) Yellow contours mark the locations of hyaline membrane in an exemplary slice. In same spatial orientation, (b, d and f) volume rendering of soft tissue (light pink), infiltrated by hyaline membrane (yellow) and lymphocytes (red) and (c, e and g) their local cell density among the lung tissue, including air compartments. Next, the segmentation of blood vessels is demonstrated for the example of a splitting blood vessel in the zoom tomogram of sample V. The segmentation was performed manually. To give an impression of the separation of a single capillary, a series of virtual slices in the xy-plane (magnified views) is shown in Figure 9a. The separation starts with the creation of a branch from the blood vessel (arrow in slice 336). In slice 326, 1.7⁢μ⁢m above slice 336, this branch evolves into an empty and separated capillary. Another 2⁢μ⁢m above, the capillary is entirely filled with cells. Further 2.3⁢μ⁢m, the capillary is empty again and has a diameter of about 2.8⁢μ⁢m. The segmentation of the blood vessel with all its separated branches is indicated for slice 336 in Figure 9b by the red lines. In this slice, three capillaries have already separated from the main vessel, while the fourth starts to emerge, indicated by the red arrow. The 3D shape of the blood vessel is illustrated by the 3D rendering of the segmentation, shown in Figure 9c. This segmentation of the blood vessel shows the potential of the datasets, which may be fully exploited in future with more advanced segmentations.
Figure 9.  Segmentation of the blood vessel network, exemplified for biopsy (V).
(a) Series of slices in z-direction illustrate the separation of a capillary (≈5μm diameter) from a blood vessel. Red arrow marks the separating capillary. In slice 314, the capillary is entirely filled with erythrocytes. (b) Red lines in the entire virtual slice (336) represent the contours of a manual segmentation of the blood vessel network. Three capillaries already separated from the main vessel, while the fourth starts to emerge, indicated by the red arrow. (c) Segmentation illustrating the 3D structure of the blood vessel network. Scale bars: 50⁢μ⁢m.