3D virtual histopathology of cardiac tissue from Covid-19 patients based on phase-contrast X-ray tomography

  1. Marius Reichardt
  2. Patrick Moller Jensen
  3. Vedrana Andersen Dahl
  4. Anders Bjorholm Dahl
  5. Maximilian Ackermann
  6. Harshit Shah
  7. Florian Länger
  8. Christopher Werlein
  9. Mark P Kuehnel
  10. Danny Jonigk  Is a corresponding author
  11. Tim Salditt  Is a corresponding author
  1. Institut für Röntgenphysik, Georg-August-Universität Göttingen, Friedrich-Hund-Platz, Germany
  2. Technical University of Denmark, Richard Petersens Plads, Denmark
  3. Institute of Anatomy and Cell Biology, University Medical Center of the Johannes Gutenberg-University Mainz, Germany
  4. Medizinische Hochschule Hannover (MHH), Germany
  5. Deutsches Zentrum für Lungenforschung (DZL), Hannover (BREATH), Germany

Abstract

For the first time, we have used phase-contrast X-ray tomography to characterize the three-dimensional (3d) structure of cardiac tissue from patients who succumbed to Covid-19. By extending conventional histopathological examination by a third dimension, the delicate pathological changes of the vascular system of severe Covid-19 progressions can be analyzed, fully quantified and compared to other types of viral myocarditis and controls. To this end, cardiac samples with a cross-section of 3.5mm were scanned at a laboratory setup as well as at a parallel beam setup at a synchrotron radiation facility the synchrotron in a parallel beam configuration. The vascular network was segmented by a deep learning architecture suitable for 3d datasets (V-net), trained by sparse manual annotations. Pathological alterations of vessels, concerning the variation of diameters and the amount of small holes, were observed, indicative of elevated occurrence of intussusceptive angiogenesis, also confirmed by high-resolution cone beam X-ray tomography and scanning electron microscopy. Furthermore, we implemented a fully automated analysis of the tissue structure in the form of shape measures based on the structure tensor. The corresponding distributions show that the histopathology of Covid-19 differs from both influenza and typical coxsackie virus myocarditis.

Editor's evaluation

In this manuscript the authors demonstrate that X-ray imaging delivers more detailed information than standard histology by analyzing 3D information in myocardial tissue obtained from COVID-19 patients. The findings are of particular interest regarding the segmentation of the vascular network and intussusceptive angiogenesis. The authors introduce the utilization of machine learning, and state-of-the-art techniques of X-ray phase contrast which is likely to advance future work in this field. Finally, with this manuscript the authors also provide new, more detailed insights into the pathologies associated with cardiac injury due to COVID-19.

https://doi.org/10.7554/eLife.71359.sa0

Introduction

The coronavirus disease 2019 (Covid-19) is caused by the severe acute respiratory syndrome coronavirus (SARS-CoV-2), predominantly entering the body via the respiratory tract. SARS-CoV-2 infects cells by binding its spike protein to the surface protein angiotensin-converting enzyme 2 (ACE2) of the host cell (Hoffmann et al., 2020). Severe cases are most frequently affected by viral pneumonia and acute respiratory distress syndrome (ARDS), with a pathophysiology distinctly different from for example influenza infection (Ackermann et al., 2020b). Mediated by a distinct inflammatory microenvironment, an uncontrolled infection can develop and result in massive tissue damage, again primarily reported in the lung. Apart from diffuse alveolar damage, the main histological hallmark of ARDS, specific findings in the lung histopathology are high prevalence of micro-thrombi and high levels of intussusceptive angiogenesis (IA) (Ackermann et al., 2020a; Ackermann et al., 2020b; Bois et al., 2021Ackermann et al., 2020c). The latter is a rapid process of intravascular septation that produces two lumens from a single vessel. It is distinct from sprouting angiogenesis because it has no necessary requirement for cell proliferation, can rapidly expand an existing capillary network, and can maintain organ function during replication (Mentzer and Konerding, 2014). The mechanistic link between branch angle remodeling and IA is the intussusceptive pillar. The pillar is a cylindrical ’column’ or ’pillar’ that is 1μm to 3μm in diameter (Ackermann and Konerding, 2015). In short, the capillary wall extends into the lumen and split a single vessel in two. Opposing capillary walls are first dilated, and intraluminal pillars form at vessel bifurcations by an intraluminal intussusception of myofibroblasts, creating a core between the two new vessels. These cells begin depositing collagen fibers into the core, providing an extracellular matrix (ECM) for the growth of the vessel lumen. The extension of the pillar along the axis of the vessel then results in vessel duplication. These structural changes of the vasculature have been reported in various non-neoplastic and neoplastic diseases (Erba et al., 2011; Albert et al., 2020, Ackermann et al., 2012). These finding underline the notion that Covid-19 is a disease driven by, and centered around, the vasculature with direct endothelial infection, thus providing SARS-CoV-2 an easy entry route into other organs, subsequently resulting in multi-organ damage (Nishiga et al., 2020; Menter et al., 2020).

Clinically, the heart appears to be a particular organ at risk in Covid-19. Acute cardiac involvement (e.g. lowered ejection fraction, arrhythmia, dyskinesia, elevated cardiac injury markers) is reported in a broad range of cases. In contrast to other respiratory viral diseases affecting the heart (e.g. coxsackie virus), in the few Covid-19 cases reported so far that included cardiac histopathology, no classic lymphocytic myocarditis characterized by a T-lymphocyte predominant infiltrate with cardiomyocyte necrosis was observed (Gauchotte et al., 2021; Kawakami et al., 2021; Tavazzi et al., 2020; Albert et al., 2020; Wenzel et al., 2020; Halushka and Vander Heide, 2021). Furthermore, the underlying pathomechanisms are still poorly understood with both direct virus induced (cellular) damage and indirect injury being discussed (Zheng et al., 2020; Wichmann et al., 2020; Gauchotte et al., 2021; Chen et al., 2020; Deng et al., 2020; Zeng et al., 2020). Particularly, it is not known to which extent the vasculature of the heart, including the smallest capillaries, are affected and whether IA is also a dominant process in this organ. More generally, one would like to delineate the morphological changes of cytoarchitecture from other well described pathologies. Recently, we have used three-dimensional (3d) virtual histology based on phase-contrast X-ray tomography as a new tool for Covid-19 pathohistology and investigated these structural changes in postmortem tissue biopsies from Covid-19 diseased lung tissue using propagation based X-ray tomography (Eckermann et al., 2020; Walsh et al., 2021). Exploiting phase contrast based on wave propagation, the 3d structure of formalin-fixed, paraffin-embedded (FFPE) tissue–the mainstay for histopathological samples worldwide- can be assessed at high resolution, that is with sub-micron voxel size and with sufficient contrast also for soft and unstained tissues (Töpperwien et al., 2018). By relaxing the resolution to voxel sizes in the range of 25 µm and stitching of different tomograms, the entire human organ can be covered and an entire FFPE tissue block ‘unlocked’ by destruction-free 3d analysis (Walsh et al., 2021).

In this work, we now focus on the 3d cytoarchitecture of cardiac tissue. We have scanned unstained, paraffin embedded tissue, prepared by a biopsy punch from paraffin-embedded tissue blocks, collected from patients which have succumbed to Covid-19 (Cov). For comparison, we have scanned tissue from influenza (Inf) and myocarditis (Myo) patients as well as from a control group (Ctr). In total, we have scanned 26 samples, all whichwihch had undergone routine histopathological assessment beforehand. We used both a synchrotron holo-tomography setup and a laboratory µCT with custom designed instrumentation and reconstruction workflow, as described in Eckermann et al., 2020. Based on the reconstructed volume data, we then determined structural parameters, such as the orientation of the cardiomyocytes and the degree of anisotropy, as well as a set of shape measures defined from the structure tensor analysis. This procedure is already well established for Murine heart models (Dejea et al., 2019). Segmentation of the vascular network enabled by deep-learning methods is used to quantify the architecture of the vasculature.

Following this introduction, we describe the methodology, which is already summarized in Figure 1. We then describe the reconstructed tissue data in terms of histopathological findings based on visual impression, and compare the different groups. We then apply automated image processing for classification and quantification of tissue pathologies. Finally, we segment the vasculature using a deep-learning-based approach based on sparse annotations and quantify the structure of the capillary network by graph representations of the segmented vessels and quantify the vasculature, both from voxel-based measures and from extracted graph representations of the segmented vessel network. From the generalized shape measures, as well as the inspection of particular vessel architectures exhibiting the IA phenomenon, distinct changes of Cov with respect to the other pathologies and to Ctr are observed. The paper closes with a short conclusions and outlook section.

Sample preparation and tomography setups.

(A) HE stain of a 3-m-thick paraffin section of one sample from a patient who died from Covid-19 (Cov-I, Scalebar: 100μm). In total, 26 postmortem heart tissue samples were investigated: 11 from Covid-19 patients, 4 from influenza patients, 5 from patients who died with myocarditis and six control samples. (B) From each of the samples, a biopsy punch with a diameter of 3.5mm was taken and transferred onto a holder for the tomography acquisition. After tomographic scans of all samples at the laboratory setup, Covid-19 and control specimens were investigated at the synchrotron. Furthermore,at the laboratory and parallel beam setup at the synchrotron, one punch with a diameter of 1mm was taken from one of the control and Covid-19 samples for investigations at high resolution. (C) Sketch of the laboratory micro-CT setup. Tomographic scans of all samples were recorded in cone beam geometry with an effective pixel size of pxeff=2μm using a liquid metal jet source (EXCILLUM, Sweden). (D) Sketch of the parallel beam setup of the GINIX endstation (P10 beamline, DESY, Hamburg). In this geometry, datasets of Covid-19 and control samples were acquired at an effective voxel size of 650nm3. One plane of each sample was covered by 3×3 tomographic recordings. For each sample a plane of 3×3 tomographic acquisitions was recorded. (E) Cone beam setup of the GINIX endstation. After the investigation in parallel geometry, the 1 mm biopsy punches of one control and Covid-19 sample were probed and a high resolution scan in cone beam geometry was recorded. This configuration is based on a coherent illumination by a wave guide and allows for high geometric magnification and effective voxel sizes below 200nm.

Materials and methods

Autopsy, clinical background, and tissue preparation

Request a detailed protocol

In total, 26 postmortem heart tissue samples were investigated: 11 from Covid-19 patients (Cov), 4 from H1N1/A influenza patients (Inf), 5 from patients who died due to coxsackie virus myocarditis (Myo), as well as 6 control samples (Ctr). The age and sex of all patients are summarized in Table 1. Detailed information about age, sex, cause of death, hospitalization, clinical, radiological, and histological characteristics of all patients is given in Appendix 2 Tab. D.1.

Table 1
Sample and medical information of patients.
Sample groupN patientsSample quantityAgeSex
Control2631 ± 72 F
Covid-19111176 ± 1310 M, 1 F
Myocarditis5543 ± 174 M, 1 F
Influenza4463 ± 93 M, 1 F

Figure 1 illustrates the sample preparation and the tomographic scan geometries used to assess the 3d cytoarchitecture on different length scales. FFPE-tissue from autopsies was prepared by standard formalin fixation and paraffin embedding. From the paraffin-embedded tissue block, sections of 3μm thickness were prepared for histomorphological assessment using conventional hematoxylin and eosin (HE) staining. One representative microscopy image of a Covid-19 patient is shown in Figure 1. An overview of HE stained sections from all samples is shown in the Appendix 1—figure 1. In previous studies, we could show the correlation of 3d X-ray phase contrast tomography data with conventional 2d histology (Eckermann et al., 2020; Töpperwien et al., 2018).

Biopsy punches with a diameter of 3.5mm were then taken and transferred onto a holder for the tomographic scans. All samples were first scanned at a laboratory µCT instrument using a liquid metal jet anode. Next, tomograms of Covid-19 and control samples were scanned at the GINIX endstation of the P10 beamline at the PETRAIII storage ring (DESY, Hamburg), using the parallel (unfocused) synchrotron beam. Finally, biopsy punches with a diameter of 1mm was taken from the 3.5mm biopsy of one control and one Covid-19 sample and scanned at high geometric magnification M using a cone beam illumination emanating from a X-ray waveguide (WG).

Tomographic recordings

Liquid metal jet (LJ) setup

Request a detailed protocol

All samples were scanned using a home-built laboratory phase-contrast µCT-setup, as sketched in Figure 1C. X-rays emitted from a liquid metal jet anode (Excillum, Sweden) are used in cone beam geometry with a geometric magnification M=x01+x12x01 controlled by the source-sample x01 and sample-detector distance x12. The spectrum of photon energy E is dominated by the characteristic Kα lines of galinstan (Ga,Zn,In alloy), in particular the Ga line EGa=9.25keV. Projections were acquired by a sCMOS detector with a pixel size of px=6.5μm coupled by a fiber-optic to a 15-m-thick Gadox-scintillator (Photonic Science, UK) (Bartels et al., 2013; Reichardt et al., 2020). In this work, data was acquired at an effective pixel size of pxeff=pxM=2μm. For each of the 1,501 angular positions 3 projections at 0.6 s acquisition time were averaged. Further, 50 flat field images before and after the tomographic acquisition, as well as 50 dark field images after the scan were recorded. The total scan time was approximately one hour per sample.

Parallel beam (PB) setup

Request a detailed protocol

All Cov and Ctr samples were also scanned with an unfocused, quasi-parallel synchrotron beam at the GINIX endstation, at a photon energy Eph of 13.8keV. Projections were recorded by a microscope detection system (Optique Peter, France) with a 50-m-thick LuAG:Ce scintillator and a 10× magnifying microscope objective onto a sCMOS sensor (pco.edge 5.5, PCO, Germany) (Frohn et al., 2020). This configuration enables a field-of-view (FOV) of 1.6mm×1.4mm, sampled at a pixel size of 650nm. The continuous scan mode of the setup allows to acquire a tomographic recording with 3000 projections over 360° in less than 2 min.3×3 tomographic acquisitions in one plane for each of the 3.5 mm biopsy punches. For each sample, one plane of the 3.5mm biopsy punch was covered by 3×3 tomographic acquisitions. Afterwards, dark field and flat field images were acquired. In total more than 150 tomographic scans (nine tomograms for each of the 17 samples) were recorded in this configuration.

Waveguide (WG) setup

Request a detailed protocol

As a proof-of-concept that subcellular resolution can also be obtained on cardiac tissue samples, a 1mm-diameter biopsy punch was taken from both a Covid-19 and control sample, both of which were previously scanned (PB geometry). The highly coherent cone beam geometry and clean wavefront of the WG illumination allows for samples to be probed at high magnification in the holographic regime. Here, the sample was aligned at M40, resulting in an effective pixel size of 159nm. Images of the Ctr were acquired by a sCMOS Camera (15μm Gadox scintillator, 2560 ×2,160 pixel) with a physical pixel size of 6.5µm (Andor Technology Ltd, UK). Cov datasets were recorded by a 1:1 fiber-coupled scintillator-based sCMOS camera (2048 × 2048 pixels, Photonic Science, Sussex, UK) with a custom 15-m-thick Gadox scintillator with pixel size of 6.5μm. For Ctr data, the photon energy was E=10keV and 1500 projections over 180 degrees were recorded with an acquisition time of 0.3 s, for the Cov sample 1500 projections were acquired for four slightly different propagation distances at E=10.8keV. The difference in acquisition time of both scans (Ctr: 10 min, Cov 3 h) is given by different wavguide channel diameters and guiding layer materials (Ctr: Ge, Cov: Si). Before and after each tomographic scan, 50 empty beam projections as well as 20 dark fields after the scan were recorded. The experimental and acquisition parameters for all imaging modalities are listed in Table 2.

Table 2
Data acquisition parameters of the laboratory and synchrotron scans.
ParameterLJ setupPB setupWG setup (Ctr/Cov)
Photon energy (keV)9.2513.810/10.8
Source-sample-dist. x01 (m)0.092900.125/0.125 0.127 0.131 0.139
Sample-detector-dist. x12 (m)0.2060.54.975
Geometric magnification M3140
Pixel size (µm)6.50.656.5
Effective pixel size (µm)20.650.159
Field-of-view h×v (mm2)4.8×3.41.6× 1.40.344×0.407/0.325× 0.325
Acquisition time (s)3× 0.60.0350.3/2.5
Number of projections150130001500
Number of flat fieldempties50100050
Number of dark field5015020

Phase retrieval and tomographic reconstruction

Request a detailed protocol

The 3d structure of the cardiac tissue was reconstructed from the raw detector images. To this end, we computed the phase information of each individual projection and performed tomographic reconstruction to access the 3d electron density distribution. For image processing and phase retrieval, we used the HoloTomoToolbox developed by our group, and made publicly available (Lohse et al., 2020)(Table 3). First, flat field empty beam and dark field corrections were performed for all raw projections. In addition, hot pixel and detector sensitivity variations were removed by local median filtering. Phase retrieval of LJ scans was carried out with the Bronnikov aided Correction (BAC) algorithm (De Witte et al., 2009; Töpperwien et al., 2018). For the PB scans, a local ring removal (width of ±40 pixel) was applied around areas where wavefront distortions from upstream window materials did not perfectly cancel out after empty beam division. Phase retrieval of PB scans was performed using the linear CTF-approach (Cloetens et al., 1999; Turner et al., 2004). Phase retrieval of WG scans was performed using a nonlinear approach of the CTF. This advanced approach does not rely on the assumption of a weakly varying phase, and iteratively minimizes the Tikhonov-functional starting from the CTF result as an initial guess. For a weakly phase-shifting sample (linear approximation) without further constraints, both approaches yield exactly the same result (Lohse et al., 2020). Apart from phase retrieval, the HoloTomoToolbox provides auxiliary functions, which help to refine the Fresnel number or to identify the tilt and shift of the axis of rotation (Lohse et al., 2020). Tomographic reconstruction of the datasets was performed by the ASTRA toolbox (van Aarle et al., 2016; van Aarle et al., 2015). For the LJ and WG scans recorded at large cone beam geometry, the FDK-function was used, while the PB was reconstructed by the iradon-function with a Ram-Lak filter.

Table 3
Phase retrieval algorithms and parameters used for the different setups.
SetupLJ setupPB setup configurationWG setup configuration
Fresnel number0.471250.00950.0017
phase retrievalBACCTFnonlinear CTF
δ/β-ratio-1/451/130
parameterα=810-3α1=10-3α1=810-4
γ= 1α2= 0.5α2= 0.2

To combine the 3×3 tomographic volumes, covering one plane of the 3.5mm biopsy in PB geometry, a non-rigid stitching tool of was used (Miettinen et al., 2019). Region-of-interest artefacts of the PB reconstructions, which led to circular low frequency artefacts at the borders of the biopsy reconstruction volume, were removed by radial fitting of cosine functions. In order to decrease the size of the stitched volume, and thus also reduce computational power needed for further analysis, the datasets were binned by a factor of 2.

Structure tensor analysis

The laboratory datasets and the stitched datasets reconstructed from the PB recordings were used for further analysis of the cardiac structure, cytoarchitecture and the corresponding pathological changes, see the workflow sketched in Figure 2. For each reconstruction of the 3d electron density map (Figure 2A), the biopsy punches were first masked based on their higher electron density compared to the surrounding air. Missing areas in the PB acquisition (from corrupted datasets) were excluded. The intensities of the reconstructions were normalized. Figure 2B shows an exemplary masked 2D slice. For each sample, the local tissue orientation and the degree of alignment was then determined from structure tensor analysis (Krause et al., 2010). Accordingly, the local structural orientation at point r can be described by a vectorw

(1) w(r)=argminv=1(I(r+v)I(r))2

with v3 and |v|=1 in voxel units. Since the vector w or set of vectors is computed from partial derivatives, one has to first compensate for the ill-posedness of computing derivatives of noisy intensity values by convolving intensities Iσ=KσI with a Gaussian kernel Kσ. The structure tensor Jρ then is defined as follows

(2) Jρ=Kρ((xIσ)2(xIσ)(yIσ)(xIσ)(zIσ)(yIσ)(xIσ)(yIσ)2(yIσ)(zIσ)(zIσ)(xIσ)(zIσ)(yIσ)(zIσ)2),

where a second convolution Kρ is applied with length scale ρ, thus defining the structural scale on which the tissue structure is analyzed/reported. Since the reconstructed electron density I(r) along a fiber is approximately constant along the fiber tangent, the vector describing the local structural orientation is given by the eigenvector with the smallest eigenvalue of the symmetric matrix Jρ. The implementation of the structure tensor analysis is provided in https://lab.compute.dtu.dk/patmjen/structure-tensor. In this work, the size of ρ, determining the sub-volume on which the structural analysis is performed, was set to 32 pixels for PB datasets and 12 pixels for LJ acquisitions. This corresponds to 20.8μm and 24μm, respectively, that is a value slightly smaller than the diameter of a cardiomyocyte (25μm). A smoothing parameter σ of 2 pixels was chosen to reduce noise. From the eigenvalues (λ1λ2λ3) of Jρ, quantitative shape measures (as first introduced for diffusion tensor MRI data) can be determined (Westin et al., 2002). These parameters describe the degree of anisotropy of the local structure orientation. Tissue structure with fiber-like symmetry are indicated by a high value of

(3) Cl=λ2-λ3λ1.
Data analysis workflow of cardiac samples.

(A) Volume rendering of a tomographic reconstruction from PB data. (B) Orthogonal slice of the masked tissue. Scale bar: 1mm (C) Shape measure distribution (Cl red, Cp green and Cs blue) of the slice shown in B. (D) Ternary plot of shape measure distribution. The peak (red) and mean (yellow) values are marked with an asterisk. (E) Overview of the training process for the neural network. (1) Random subvolumes (containing labeled voxels) are sampled from the full volume and are collected in a batch. (2) The batch is fed through the neural network, resulting in (3) a segmentation (top) and labels for one subvolume (bottom). (4) The dice loss is computed from segmented subvolumes based on labeled voxels, and the parameters of the neural network are updated. (F) Scheme of branching and the relation to degree of the vessel nodes obtained by a graph representation of the segmented microvasculature.

Plane-like (lamellar) symmetry is described by a high value of

(4) Cp=λ1-λ2λ1,

and isotropic structures are described by a high value of the spherical shape measure

(5) Cs=λ3λ1.

The shape measure distribution of the exemplary slice is shown in Figure 2C. Red areas indicate a high Cl value and correlate with the well aligned chains of cardiomyocytes. Planar structures as collagen sheets and separated muscle bundles show a high Cp value and are color-coded in green. Isotropic areas as blood filled vessels are represented by a high Cs value (blue). The values of the three measures range between zero and one, and sum up to one

(6) Cl+Cp+Cs=1.

Thus, one of the three shape measures is redundant. The data can be plotted in a ternary diagram as used to represent phase diagrams of ternary mixtures (see Figure 2D). To characterize the distribution of the shape measures for each sample, a principal component analysis (PCA) was performed. Note, that for the LJ datasets, the paraffin surrounding the cardiac tissue was removed by an intensity-based mask. Since one axis of the shape measure is redundant, the distribution of all data points can be described by two eigenvectors (u1,u2 with the largest eigenvalues (η1,η2)). The PCA analysis is equivalent to a two-dimensional Gaussian with standard deviation η1,η2. The two eigenvectors (u1,u2) can be represented by the major and minor axis of an ellipse centred around the mean (μl,μp,μs) (yellow asteroid) representing the ‘point cloud’ of all shape measures. The eccentricity of the ellipse is given by

(7) e=1-η2η1

and describes how much the ellipse deviates from being circular. The area of the ellipse is given by Aη=πη1η2 and is a measure for the dispersion of the shape measure distribution. The eccentricity indicates whether the dispersion is isotropic in the plane of the shape parameters. Large values of e indicate a sharp elongated distribution along the major axis of the ellipse.

Segmentation by deep learning

Request a detailed protocol

A deep learning approach based on the V-Net architecture (Milletari et al., 2016) was used to segment the vascular network in the PB datasets. The V-Net can be regarded as a 3D version of the popular U-Net architecture (Ronneberger et al., 2015) often used for segmentation of medical images. Training was performed using the Dice loss (Milletari et al., 2016) and the ADAM optimizer (Kingma and Ba, 2015) with step size 0.001 and hyperparameters β1=0.9 and β2=0.999. To avoid the need of a fully labeled training dataset, a training strategy using sparsely annotated data sets was adopted, inspired by Çiçek et al., 2016. In each dataset, a small number of axis-aligned 2D slices was annotated manually, and the Dice loss was evaluated only for these annotated voxels. Prior to training, the annotated volumes were split into a training set and a smaller validation set. The network was trained on the training set, while the quality of the current model (network weights) was tested on the validation set, as sketched in Figure 2E. Instead of segmenting the entire volume before computing the loss, batches of 12 random subvolumes of size 96 × 96 × 96 voxels were selected, ensuring that each contained annotated voxels. These were then fed into the network, the loss was computed, and the parameters (network weights) were updated. After running on 256 subvolumes, the network was evaluated by running it on the validation set. Rotations by 90 degrees and mirror reflections (axis flips) were used both on the training and the validation subvolumes to augment the data. The neural network code of this implementation was uploaded to GitHub (github.com/patmjen/blood-vessel-segmentation; Jensen, 2021 copy archived at swh:1:rev:783df24c3068e35f2ae994cab095b4318c755b29).

A separate model was trained for a Covid-19 volume (Cov-IV) and a control volume (Ctr-III). The models were trained for 24 hr (~900 epochs) using an NVIDIA Tesla V100 32 GB GPU, and the model version which achieved the highest validation score during the training was kept. Finally, the training was performed over two rounds. First, an initial training and validation set was created to train the model. Then, the training set was improved by adding additional annotations to areas which were falsely segmented, and a new model was trained on the improved data.

As the segmentation masks produced by the neural networks typically contained a number of errors, a post-processing pipeline was designed to reduce the errors’ effect. The first step is to reduce the number of false positives. These typically materialize as small, roughly spherical regions of background which was erroneously detected as blood vessels. To remove them, the structure tensor shape measures Cl, Cp, and Cs are computed for the segmentation mask (treating background as 0 and foreground as 1) with σ and ρ set to 1 and 8 voxels, respectively. Then, all connected components with a volume less than 104 voxels or a mean value of Cs greater than 0.2 are removed. The thresholding on Cs ensures that isotropic components are removed regardless of their size while still preserving smaller sections of correctly segmented blood vessels. The last step is to reduce the number of false negatives by reconnecting segments of blood vessels which are disconnected due to small errors in the segmentation. Since endpoints of blood vessels will typically have a large value of Cs, small gaps in the vessels can be closed by performing a morphological closing of the isotropic regions of the segmentation mask. Specifically, the cleaned binary mask, B^, is given by

(8) B^=max(B,close(ClB,S4)>0.2) ,

where B is the original binary mask (after the first post processing step), Cl is the line-like measure for all voxels in B, and close(ClB,S4) denotes a closing of the elementwise product between Cl and B with a ball of radius 4. For performance reasons the closing uses an approximated ball as described in Jensen et al., 2019.

Quantification of the vascular system

Request a detailed protocol

A quantitative description of the vascular system was achieved by modeling the segmented vessels as a mathematical graph. A graph consists of a set of vertices and a set of edges where each edge connects a pair of vertices. If vertices are connected via an edge they are said to be neighbors and the degree of a vertex (nodes) n is equal to its number of neighbors. In Figure 2F a sketch of a vessel graph is shown for a straight vessel and for a vessel with multiple branching points. The degree of connectivity n is added to the sketch. This gives a natural correspondence to the complex vascular system by modeling bifurcation points as vertices and the blood vessels between pairs of bifurcation points as edges. Furthermore, structural phenomena such as excessive vessel bifurcation and intussusceptive angiogenesis can now be detected by, respectively, a large number of high degree vertices and loops in the graph. The graph corresponding to the vascular system is extracted from the segmentation created by the neural network. First, a skeletonization (Lee et al., 1994) is computed, which reduces all structures in the binary volume to 1-voxel wide centerlines without changing the connectivity. These centerlines are then converted to a graph as described in Kollmannsberger et al., 2017. Once the graph is constructed the vertex degrees can readily be extracted by counting the number of edges connected to each vertex. Loops are detected using the algorithm from Gashler and Martinez, 2012 which detects all atomic cycles in a given graph. A cycle is a path through the graph that begins and ends at the same vertex without reusing edges. An atomic cycle is a cycle which cannot be decomposed into shorter cycles. Only reporting atomic cycles is more robust, since small errors in the segmentation may cause the skeletonization to contain long cycles that do not correspond to anatomical structures. The 3d data sets (including tomographic reconstructions and segmentations) was visualized using the software Avizo (Thermo Fisher Scientific).

Vascular corrosion casting, scanning electron microscopy, and morphometry

Request a detailed protocol

The microvascular architecture of Covid-19 hearts was also examined using scanning electron microscopy (SEM) and microvascular corrosion casting. So far, corrosion casting coupled with SEM represents the gold standard for assessing the subtypes of angiogenesis. The afferent vessels of heart specimens were cannulated with an olive-tipped cannula. The vasculature was flushed with saline (at body temperature) followed by glutaraldehyde fixation solution (2.5%, pH 7.4, Sigma Aldrich, Munich, Germany). Fixation was followed by injection of prepolymerized PU4ii resin (VasQtec, Zurich, Switzerland) mixed with a hardener (40% solvent) and blue dye as casting medium. After curing of the resin, the heart tissue was macerated in 10% KOH (Fluka, Neu-Ulm, Germany) at 40°C for 2–3 days. Specimens were then rinsed with water and frozen in distilled water. The casts were freeze-dried and sputtered with gold in an argon atmosphere and examined using a Philips ESEM XL-30 scanning electron microscope (Philips, Eindhoven, Netherlands). Vascular morphometry of variants of angiogenesis were then assessed: high-power images of the capillary network were scanned and quantified.

Results and discussion

Reconstructed electron density: laboratory data

Figure 3 shows representative slices of the tomographic reconstruction for all samples scanned at the laboratory LJ setup. The image quality is sufficient to identify the cytoarchitecture and main structural features of interest, such as the general orientation of the cardiomyocytes, large arteries and veins, as well as smaller capillaries. Occasionally, artefacts from sample preparation, such as small air filled micro-fractures of the paraffin, also appear in the reconstructions. In the top row of Figure 3, two annotated slices representative for the Covid-19 and control group are shown enlarged. In the following, the structural appearance of the different groups (Ctr, Cov, Inf and Myo) is briefly described.

Overview of reconstruction volumes: Laboratory setup.

For each sample analyzed at the LJ µ-CT setup one slice of the reconstructed volume is shown. In the top row, a slice of a tomographic reconstruction of a control sample (Ctr-I) and of a sample from a patient who died from Covid-19 (Cov-I) are shown. Below, further slices from control (Ctr-II to Ctr-VI), Covid-19 (Cov-II to Cov-XI) as well as myocarditis (Myo-I to Myo-V) and influenza (Inf-I to Inf-IV) samples are shown. Scale bars: 1mm.

Control (Ctr)

The reconstructions of the control hearts are shown in the top row (Figure 3 (Ctr-I to Ctr-VI)). Biopsies Ctr-I to Ctr-III and Ctr-IV to Ctr-VI were taken from different areas of the same heart, respectively. In general, the cardiac structure with interload cardiomyocytes and vasculature of the control group is well preserved. The cardiomyocytes are arranged in close proximity and form bundled elongated myocyte chains. Vessels appear as bright tubes within the dense, homogeneous muscle tissue and only a few blood residues can be found in the vessels. Ctr-III differs from the other control samples. The alignment of the cardiomyocytes is not directed along the same direction, and the amount of collagen sheets and paraffin inclusions is comparably high. Further, a high amount of adipose tissue can be identified, as accumulations of less electron-dense (i.e. brighter) spheroids, see for example the top of the slice. Ctr-III also shows a high amount of collagen sheets, which appear as dark stripes in the reconstructions. Ctr-V contains many electron-dense spheres.

Covid-19 (Cov)

The cardiac samples of the hearts from patients who died from Covid-19 are shown in the next two rows of Figure 3 (Cov-I to Cov-XI). Compared to Ctr, all Cov samples show a high amount of blood filled, ectatic vessels with abrupt changes in diameter, plausibly correlating to micro-thrombi. The cardiomyocytes are not densely packed with substantial interstitial edema, and correspondingly there is a high amount of paraffin inclusions between the cells. This may also explain a higher amount of micro-fractures (e.g. in Cov-I and IV) in the paraffin, which are filled with air. Furthermore, Cov-I also shows an inflammatory infiltrate, predominantly consisting of macrophages, around the intramyocardial vessel, marked in the corresponding slice (top, right) of Figure 3.

Coxsackie virus myocarditis (Myo)

In Figure 3, representative slices from tomographic reconstructions of biopsies of patients who died from coxsackie myocarditis (Myo-I to Myo-V) are shown. The tissue of the Myo group is almost as densely packed as the Ctr group. Only in the biopsy of Myo-III, which was sampled near an artery, some large paraffin inclusions between the cardiomyocytes are visible. Characteristic for all myocardits samples is a high amount of lymphocytes, which appear as small electron-dense spheres in the reconstructions. They are primarily located close to vessels (as in Myo-II), but also appear inside the ECM between cardiomyocytes (Myo-I), or infiltrate extensive areas of tissue devoid of vital cardiomyocytes, corresponding to necrosis (Myo-V).

Influenza (Inf)

The biopsies taken from patients who succumbed to H1N1/A influenza (Inf-I to Inf-IV) are shown in the bottom row of Figure 3. The tissue structure in this group is also densely packed. Inf-IV shows a high amount of blood filled vessels with abrupt changes in caliber, plausibly correlating to micro-thrombi. Otherwise, changes include lymphocytic infiltration and regions devoid of vital cardiomyocytes indicating necrosis, similar to (Myo).

In summary, the quality of the reconstructions from laboratory data was already sufficiently high to identify the main anatomical features of the cardiac tissue, readily by eye in selected slices. The full reconstruction volumes were therefore targeted by automated geometric analysis based on a structure tensor approach, as described in the next section. However, smaller capillaries and subcellular features were not resolved in the laboratory LJ setup configuration. Thus, imaging using high coherent synchrotron radiation was chosen to analyze vascular changes within the tissue.

Reconstructed electron density: synchrotron data

PB setup configuration

The samples from Ctr and Cov patients were scanned in the PB setup configuration of the GINIX endstation (Hamburg, DESY). Compared to the laboratory acquisitions, this allowed for smaller effective voxel sizes and enabled a higher contrast for smaller tissue structures as erythrocytes and capillaries (as shown in AppendixC Fig. C). Slices of the tomographic reconstruction of the 3d electron density distribution are shown in AppendixC Fig. C and were used for the segmentation of the vascular system.

WG setup

In order to further explore high-resolution imaging capabilities, tomograms of two selected biopsies (Ctr-VI and Cov-III) with a diameter of 1mm were recorded at the WG setup of the GINIX endstation, exploiting cone beam magnification and high coherence filtering based on the waveguide modes. Figure 4 shows the corresponding results. A cut of the entire control volume with a size of about 340 × 340 × 400 µm3 is shown in Figure 4. Figure 4 shows a slice through the tomographic reconstruction perpendicular to the orientation of the cardiomyocytes. A closer inspection of a single cardiomyocyte marked with a red box is shown on the right. The nucleus of the cell with nucleoli can be clearly seen. Within the cytosol, the myofibrils appear as small discs in the slice. Figure 4C shows a second slice through the 3d volume which is oriented along the orientation of the cardiomyocytes. In this view, intercalated discs can be identified. They appear as dark lines connecting two cardiomyocytes. A magnification of the area is marked with a red box. In this view, the myofibrils can be identified as elongated lines within the cell. This region also contains a nucleus of one cardiomyocyte, but also an intercalated disc at the bottom of the image. The tomographic reconstruction of the Cov sample is shown in the lower part of Figure 4 in the same manner as the Ctr. In this dataset capillaries, nuclei and myofibrils can also be identified. The volume contains smaller capillaries compared to the control, but this circumstance is probably due to a different location within the myocardium. The most important difference between the Ctr and Cov sample is the presence of small bars in the lumen of capillaries in the Cov sample. These intraluminal pillars are an indicator for IA.

High-resolution tomogram of cardiac tissue recorded in cone beam geometry.

(A) Volume rendering of a tomographic reconstruction from a control sample recorded in cone beam geometry based on a wave guide illumination. After the analysis in parallel beam geometry, a biopsy with a diameter of 1mm was taken from the 3.5mm biopsy punch. This configuration revealed sub-cellular structures such as nuclei of one cardiomyocytes, myofibrils and intercalated discs. (B) Slice of the reconstructed volume perpendicular to the orientation of the cardiomyocytes. The red box marks an area which is magnified and shown on the right. One cardiomyocyte is located in the center of the magnified area. In this view, the nucleus can be identified. It contains two nucleoli, which can be identified as dark spots. The myofibrils appear as round discs. (C) Orthogonal slice which oriented along the orientation of the cardiomyocytes. A magnification of the area marked with a red box. In this view, a nucleus but also the myofibrils can be identified as dark, elongated structures in the cell. Further, an intercalated disc is located at the bottom of the area. (D) Volume rendering of a tomographic reconstruction from a Covid-19 sample. Slices orthogonal (E) and along (F) to the cardiomyocyte orientation are shown on the right. In the magnified areas, a nucleus of an endothelial cell and an intraluminar pillar -the morphological hallmark of intussusceptive angiogenesis- are visible. Scale bars: orthoslices 50μm; magnified areas 10μm.

Since the FOV in this configuration is limited, and stitching of larger volumes required more beamtime than available, quantitative and statistical analysis was performed only on the datasets acquired in the laboratory and in PB geometry. At the same time, this proof-of-concept shows that much more structural information could be exploited by stitching tomography and speeding-up the measurement sequence in the WG configuration.

The tomographic datasets recorded at the WG setup in WG configuration as well as the PB datasets used for the segmentation of the vascular system and the respective laboratory datasets were uploaded to https://doi.org/10.5281/zenodo.5658380 (Reichardt et al., 2021).

Automated tissue analysis and classification of pathologies

Next, the reconstructed 3d tissue structure is analyzed by an automated workflow involving differential operators and subsequent statistical representations based on the structure tensor analysis. Instead of semantic analysis of specific structures (vessels, cardiomyocytes, ect), which is considered further below, we first target geometric properties encoded by gray value derivatives, possible prototypical distribution of these parameters in a sample, and the respective variations within and between groups. This can then later be interpreted also in view of semantic image information. A high local anisotropy and consistent orientation field, for example, can be indicative of an intact tissue with well-ordered cardiomyocyte chains. For all samples, eigenvector and eigenvalues were computed for all sampling points in the reconstructed volume. This information then includes the orientation (quasi-)vector as defined by the smallest eigenvector, as well as the shape measures for all points. As a word of caution, however, one has to keep in mind that these properties also depend on tissue preservation and preparations, as well as on the measurement and reconstruction. For this reason, the latter has to be carried out using identical workflows and parameters for all samples.

Figure 5 shows the results of the structure tensor analysis for all samples reconstructed from LJ scans. In Figure 5A the mean values of the shape measures (μl,μp,μs) for all datasets are plotted in a shape-measure diagram, constructed as for ternary mixtures. Sample groups are indicated by color: control-green, Covid-19-red, myocarditis-blue and influenza-yellow. Already in this plot, differences between the groups can be identified. Compared to the Ctr, the pathological groups are shifted towards lower Cl, indicating a less-pronounced fiber-like structure, and to higher Cs, reflecting a larger amount of isotropic symmetry. The Cov, Inf and Myo groups differ mainly in the Cp coefficient. From Inf, to Myo and Cov, the point clouds of each group exhibit successive shifts toward increased Cp. However, these differences in µ are quite small, and it is not possible to classify samples only based on the average value of the shape measure. Instead, the distribution of real-space sampling points in each sample should be taken into account. Figure 5B and C show the area Aη and the eccentricity e, respectively, of the ellipse formed by the PCA eigenvectors u1,u2, for each sample, color-coded by groups. The corresponding box-whisker plots indicate a significant difference in Aη between Cov and Ctr (Welch t-test, p=0.0389) as well as a Cov and Inf (Welch t-test, p=0.0403). Concerning e, Cov tissues differs also from Myo (Welch t-test, p=0.0611). Small values of Aη, as obtained for Ctr, indicate a homogeneous tissue structure, while large values are obtained for samples with a more heterogeneous tissue composition. The parameters for each sample group are tabulated in Table 4. The large intra-group variance reflects the pronounced variability between individual subjects, which is in line with experience of conventional histology. The complete summary of all samples individually is given in Appendix 1—figure 1, Appendix 2—table 2. The results for the stitched tomographic datasets (PB setup) of Cov and Ctr are also shown in Appendix 1—figure 1.

Table 4
Parameters of the cardiac tissue obtained from LJ reconstructions.

For all sample groups the mean value and standard deviation of the mean shape measures μl¯, μp¯, μs¯ area of the elliptical fit Aη¯ (%) and the eccentricity e¯ is shown.

Groupμl¯μp¯μs¯Aη¯(%)e¯
Control0.60± 0.110.18± 0.070.22± 0.0611.98± 6.420.61± 0.13
Covid-190.44±0.120.23±0.030.32±0.1116.92± 2.910.61± 0.09
Myocarditis0.47±0.140.21± 0.020.33±0.1316.69± 5.060.51± 0.12
Influenza0.49±0.110.16±0.020.35±0.1213.44± 1.310.63± 0.07
Clustering of LJ data sets.

(A) Ternary diagram of the mean value of the shape measures for all datasets. The control samples (green) show low Cs values, while samples from Covid-19 (red), influenza and myocarditis (blue) patients show a larger variance for Cs. (B) The fitted area of the elliptical fit from the PCA analysis of the shape measure distribution is an indicator for the variance in tissue structure. For Control and influenza sample this value differs significantly from the Covid-19 tissue. (C) The eccentricity of the fit indicates if the structural distribution in shape measure space has a preferred direction along any axis. The value of the myocarditis samples is comparable low.

Characterization of the vascular system

Figure 6 reports on the segmentation and analysis of the vasculature. A surface rendering of the segmented vessels is shown in the top row, on the left for a Ctr sample (Ctr-III) and on the right for a Cov sample (Cov-IV). In Ctr, the vessels are well oriented and show a relatively constant diameter and a smooth surface. In Cov, the vessels show large deviations in diameter and the surface of the vessels is not as smooth as in Ctr. Furthermore, closed loops within the microvasculature can be identified. In Figure 6C, one of these vessel loops (marked with a blue line) in the Cov dataset is highlighted by a minimum intensity projection over ±30 slices around the centered slice. This pathological formation of a loop is indicative for an intermediate state in the process of IA. The corresponding vessel segmentation is depicted in Figure 6D, with a simplified vessel graph superimposed as black lines. Based on the simplified vessel graph, the connectivity of the capillaries can further be quantified. In total 19,893 nodes for the Cov sample and 8068 nodes in the segmentation of the Ctr were used. Figure 6E shows the probability density function (PDF) of the degree of connectivity n for control and Covid-19 samples. It indicates a higher amount of branching points in the Covid-19 sample. This is also confirmed by the ratio of endpoints of vessels (n=1) to the branching points (n3). Note, that the amount of nodes with n>3 is almost negligible. While the Ctr data shows approximately the same number of endpoints and branching points, the Cov segmentation show almost a ratio of 1:1.5, indicating a higher degree of cross-linking or loop formation of the capillary network.

Segmentation of the vascular system in cardiac samples.

(A) Segmentation of the vessels of a Ctr sample. The vessels are well oriented and show a relatively constant diameter. (B) Segmentation of the vessels of a Covid-19 sample. The vessels show large deviations in diameter and the surface of the vessels is not as smooth as in the control sample. (C) Filtered minimum projection of an area of the reconstructed electron density of the Cov sample to highlight a vessel loop marked in blue. (D) Surface rendering of the segmented vessel and vessel graph in an area of the Cov sample. Scale bars 25μm. (E) Comparison of node degree n between control and Covid-19. Ratio refers to the number of graph branch points (n > 2) divided by the number of end points (n = 1). (F) Exemplary scanning electron microscopy image of a microvascular corrosion casting from a Covid-19 sample. The black arrows mark the occurrence of some tiny holes indicating intraluminar pillars with a diameter of 2μm to 5μm, indicating intussusceptive angiogenesis. Magnification x800, scale bar 20μm.

An exemplary scanning electron micrograph of a Covid-19 sample is shown in Figure 6F. IA was identified via the occurrence of tiny holes with a diameter of 2–5µm in SEM of microvascular corrosion casts. Capillaries display the presence of characteristic intussusceptive pillars (marked by black arrows).

Summary, conclusion, and outlook

This is the first report of a comprehensive 3d analysis of cardiac involvement in tissue of Covid-19, influenza and coxsackie virus infections using X-ray phase-contrast tomography of human FFPE heart tissue. In summary, a high amount of distinct caliber changes of blood filled capillaries in samples of Covid-19 (Cov) patients was identified compared to the control group (Ctr) as well as to coxsackie virus myocarditis (Myo) and influenza (Inv). This can readily be explained by a much higher prevalence of micro-thrombi in Cov compared to other viral pneumoniae (e.g. influenza), as has previously been reported in Covid-19 lungs. Most importantly, high-resolution synchrotron data revealed distinct alterations of the vasculature, with larger variation in vessels diameters, intravascular pillars and amount of small holes, indicative for IA. Branching points of vessels were quantified based on graph representations, after segmentation of vessels based on deep learning. For this purpose, a network for 3d datasets (V-net) was trained with sparse annotations. In Cov, the vasculature also showed a higher degree of branching. Further, SEM data showed a high amount of holes in the capillaries, indicating the presence of multiple intussusceptive pillars as a first stage of IA. The presence of intraluminar pillars was also confirmed by the high resolution reconstruction obtained from WG acquisitions. Accordingly, we could -for the first time– visualize the presence of IA via destruction-free X-ray phase-contrast tomography not only in the heart but also for the first time in FFPE-tissue. Thus, IA is also a hallmark of Covid-19 inflammation in the heart, analogous to pulmonary previously reported for lung (Ackermann et al., 2020b). This finding is in line with the concept of Covid-19 as a systemic and multi-organ angiocentric entity.

The reconstructed electron density of the Cov sample group also showed that concordant with the edema found in conventional histopathology assessment, the cardiomyocytes are not as densely packed as in the control (Ctr) group, leading to larger paraffin inclusions between the cells. Pathological alterations of the tissue architecture were further quantified in terms of non-semantic shape measures, derived from gray value differential operators, using the structure tensor approach. Since the shape measures not only depend on the tissue structure but also on the data acquisition and reconstruction parameters, the entire data acquisition and workflow was optimized and then kept constant for the entire sample series, covering the different pathologies (Cov, Inf, Myo) and control (Ctr) group samples. Importantly, this was already possible at a home-built compact laboratory µCT, based on a liquid metal jet source and optimized phase retrieval, which is important for future translation and dissemination of the methodology developed here. Fully automated PCA analysis then yielded the eigenvectors of the structure tensor at each sampling point of the reconstruction volume, and for each sample. The corresponding distributions showed significant difference in architecture between Cov from all other groups Inf, Myo or Ctr groups, and these differences could be interpreted again by inspection of the reconstruction volumes, that is reflecting for example tissue compactness, orientation of the cardiomyocytes and the degree of anisotropy.

Compared to related studies (Walsh et al., 2021), which focused on the analysis of entire human organs, we investigated the cardiac structure from the scale of 3.5 mm biopsy punches down to a resolution showing subcellular and supramolecular structures such as myofibrils and intussusceptive pillars.

Future improvements in segmentation and quantification will be required to fully exploit the structural data acquired here, or in similar studies. To this end, augmented image processing algorithms, deep learning, classification for example based on optimal transport, and the consolidation of the above in form of specialized software packages has to be considered. Technical improvements towards higher resolution and throughput can also be foreseen. Already at present, parallel beam synchrotron data acquisition (GINIX endstation, P10 beamline of PETRA III/DESY) completes a biopsy punch tomogram within 1.5min, at a pixel size of 650nm, and a volume throughput of 107μm3/s. Importantly, the image resolution and quality is sufficient to segment vasculature and cytoarchitectural features of interest, also and especially for standard unstained paraffin-embedded tissue used in routine diagnostics. The data acquisition rate and dwell time in the range of 10ms to 20ms (per projection) is dictated by detector readout, motor synchronisation, and data flow rather than by photon flux density for the PB setup. This is also underlined by the fact that (single-crystal) attenuators had to be used to prevent detector saturation. The situation is entirely different, however, for the waveguide cone beam setup, where the lower waveguide exit flux density, which comes with the significantly higher coherence and resolution, requires acquisition times of 200ms to 2500ms. Here, the projected source upgrade foreseen for PETRA IV will provide a significant gain in resolution and throughput. Robotic sample exchange will therefore be required, as well as a serious upscaling of the data management and online reconstruction pipeline. First reconstructions of heart biopsies exploiting the enhanced coherence and resolution of a waveguide holo-tomography setup already indicate that this is a very promising direction. With our presented workflow, especially in view of the laboratory system, we have for the first time implemented destruction free analysis of the ubiquitous FFPE embedded tissue readily available in every pathology lab around the world, based on an automated structure tensor and shape measures. This represents a first and major step in unlocking the extensive international FFPE archives for sub-light-microscope resolution destruction-free 3d-tissue analysis, unfolding manifold future research possibilities in human diseases far beyond Covid-19. This approach has been successfully used to classify the distinct changes in the myocardial cytoarchitecture induced by Covid-19. More importantly still, we have provided first proof for the suspected presence of IA in cardiac Covid-19 involvement, putting forward morphological evidence of a so far imprecisely defined clinical entity of great importance.

Appendix 1

Supplementary figures

Appendix 1—figure 1
HE stain of all cardiac samples .

Scale bar: 5mm.

Appendix 1—figure 2
Reconstructions of the LJ compared to the PB setup.

Comparison of the data quality of laboratory and synchrotron measurements. (A) slice of a laboratory reconstruction at a voxelsize of 2μm. A region of interest containing a branching vessel is marked by a blue box which is shown in (B). The same area cropped from a tomographic reconstruction at the PB setup at a voxelsize of 650nm is shown in (C). The smaller voxelsize, higher contrast and SNR of the PB scans is necessary to segment the vascular system. Scale bars: (A) 1mm, (B,C).50μm

Appendix 1—figure 3
Shape measure of all Covid-19 and control samples reconstructed from PB data.

Slices of the reconstructed electron density (stitched volumes of 3 ×three tomographic reconstructions), the corresponding slice of the shape measure and the ternary plot of the shape distribution in the entire volume are shown. Corrupted datasets were excluded from the analysis and masked in white. Scale bar: 1mm.

Appendix 2

Supplementary Information: Medical background and datasets

Medical Information

Appendix 2—table 1
Sample and medical information.

Age and sex, clinical presentation with hospitalization and treatment. RF:respiratory failure, CRF: cardiorespiratory failure, MOF: multi-organ failure, V: ventilation, S: Smoker, D: Diabetes TypeII, H: Hypertension, I: imunsupression

Sample no.Age, sexHospitalization (days), clinical, radiological and histological characteristics
Cov-I86,M5d, RF, D, H, I
Cov-II96,M3d, RF, H
Cov-III78,M3d, CRF, V, D, S, H
Cov-IV66,M9d, RF, V, S, H
Cov-V74,M3d, RF, D, S, H
Cov-VI81,F4d, RF, S, H
Cov-VII71,M0d, V
Cov-VIII88,M2d, V, H, I
Cov-IX85,M5d, V, S, H
Cov-X58,M7d, V, H
Cov-XI54,M15d, V
Ctr-I to Ctr-III26, F-
Ctr-IV to Ctr-VI36, F-
Myo-I57,MV, H
Myo-II23,M
Myo-III59,MS, H, D
Myo-IV50,MV, S, D
Myo-V25,F
Inf-I74,M9d, CRF into MOF, V, S, H
Inf-II66,F17d, MOF, V, H
Inf-III56,M3d, CRF into MOF, V
Inf-IV55,M24d, RF into MOF, V, S

Structural analysis

Appendix 2—table 2
Parameters of the cardiac tissue (laboratory data).
SampleMean (Cl,Cp, Cs)Fitted areaEccentricity
Ctr-I(0.6508, 0.1069, 0.2423 )7.31940.5607
Ctr-II( 0.5167, 0.1907, 0.2926 )11.51300.5736
Ctr-III( 0.5074, 0.2427, 0.2499)23.74430.4128
Ctr-IV( 0.7434, 0.1166, 0.1400 )5.90260.6757
Ctr-V( 0.7038, 0.1495, 0.1467 )9.57630.7896
Ctr-VI( 0.4765, 0.2835, 0.2400 )13.79730.6688
mean(0.60 ± 0.11. 0.18 ± 0.07, 0.22 ± 0.06)11.98 ± 6.420.61 ± 0.13
Cov-I( 0.5398, 0.2327, 0.2275)12.70520.6696
Cov-II( 0.4676, 0.2550, 0.2774 )17.03470.6059
Cov-III( 0.5896, 0.2526, 0.1578)11.88450.7399
Cov-IV( 0.5911, 0.1833, 0.2255 )16.30400.6765
Cov-V( 0.3371, 0.2505, 0.4124)16.34450.4081
Cov-VI( 0.5184, 0.2279, 0.2537)19.19540.6044
Cov-VII(0.3912, 0.2262, 0.3826)19.82060.6530
Cov-VIII( 0.5227, 0.1776, 0.2997)15.07910.6033
Cov-IV(0.3253, 0.2851, 0.3897 )20.57680.5329
Cov-X(0.3283, 0.2446, 0.4271 )16.99890.6266
Cov-XI( 0.2484, 0.2314, 0.5202 )20.18150.5407
mean(0.44±0.12,0.23±0.03,0.32±0.11 )16.92 ± 2.910.61 ± 0.09
Myo-I(0.5777, 0.2018, 0.2206 )9.55280.4656
Myo-II(0.3887, 0.1943, 0.4170 )13.78530.4899
Myo-III( 0.5984, 0.2081, 0.1935 )22.47680.6202
Myo-IV( 0.4974, 0.1908, 0.3117 )18.3306?0.6149
Myo-V(0.2664, 0.2402, 0.4933 )19.32120.3689
mean(0.27 ± 0.14, 0.24 ± 0.02, 0.49 ± 0.13)16.69 ± 5.060.51 ± 0.12
Inf-I(0.3561, 0.1714, 0.4724 )14.93930.6808
Inf-II( 0.4423, 0.1376, 0.4201 )11.7445?0.5991
Inf-III( 0.6150, 0.1361, 0.2489 )13.59880.7198
Inf-IV( 0.5404, 0.1849, 0.2747 )13.48850.5561
mean(0.49 ± 0.11, 0.16 ± 0.02, 0.35 ± 0.11)13.44 ± 1.310.63 ± 0.07

Datasets

The tomographic datasets recorded at thein WG setup as well as the PB datasets used for the segmentation of the vascular system were uploaded to https://doi.org/10.5281/zenodo.5658380.

Data availability

The tomographic datasets recorded in WG configuration as well as the PB datasets used for the segmentation of the vascular system and the respective laboratory datasets were uploaded to https://doi.org/10.5281/zenodo.4905971. Additional data (raw data, PB and laboratory reconstructions, structure tensor analysis) is curated here at University of Göttingen and at DESY can be obtained upon request from the corresponding author (tsaldit@gwdg.de); due to the extremely large size >15TB it cannot presently be uploaded easily to a public repository. The implementation of the structure tensor analysis is provided in https://lab.compute.dtu.dk/patmjen/structure-tensor. The neural network code used for the segmentation of the vasculature was uploaded to GitHub (https://github.com/patmjen/blood-vessel-segmentation copy archived at https://archive.softwareheritage.org/swh:1:rev:783df24c3068e35f2ae994cab095b4318c755b29).

The following data sets were generated
    1. Reichardt M
    2. Jensen PM
    3. Dahl VA
    4. Dahl AB
    5. Ackermann M
    6. Shah S
    7. Länger F
    8. Werlein C
    9. Kühnel M
    10. Jonigk D
    11. Salditt T
    (2021) Zenodo
    3D virtual Histopathology of Cardiac Tissue from Covid-19 Patients based on Phase-Contrast X-ray Tomography.
    https://doi.org/10.5281/zenodo.4905971

References

  1. Conference
    1. Çiçek Ö
    2. Abdulkadir A
    3. Lienkamp SS
    4. Brox T
    5. Ronneberger O
    (2016)
    In International conference on medical image computing and computer-assisted intervention
    3d u-net: learning dense volumetric segmentation from sparse annotation.
  2. Conference
    1. Jensen PM
    2. Trinderup CH
    3. Dahl AB
    4. Dahl VB
    (2019) In Scandinavian Conference on Image Analysis
    Zonohedral approximation of spherical structuring element for volumetric morphology.
    https://doi.org/10.1007/978-3-030-20205-7
  3. Conference
    1. Kingma DP
    2. Ba J
    (2015)
    In International Conference on Learning Representations, ICLR
    Adam: A method for stochastic optimization.
  4. Conference
    1. Milletari F
    2. Navab N
    3. Ahmadi SA
    (2016) IEEE
    V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation.
    https://doi.org/10.1109/3DV.2016.79
  5. Conference
    1. Ronneberger O
    2. Fischer P
    3. Brox T
    (2015)
    In International Conference on Medical image computing and computer-assisted intervention
    U-net: Convolutional networks for biomedical image segmentation.

Decision letter

  1. Hina Chaudhry
    Reviewing Editor; Harvard University, United States
  2. Matthias Barton
    Senior Editor; University of Zurich, Switzerland

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "3D virtual Histopathology of Cardiac Tissue from Covid-19 Patients based on Phase-Contrast X-ray Tomography" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Guest Editor and a Senior Editor. The reviewers have opted to remain anonymous.

The reviewers have discussed their reviews with one another, and the Reviewing Guest Editor has drafted this to help you prepare a revised submission.

Essential revisions:

Please refer to revisions requested by the reviewers below, particularly as outlined by Reviewer 1 whose critique will greatly help to strengthen this manuscript and its conclusions.

Reviewer #1 (Recommendations for the authors):

Below the authors will find my additional comments to increase the clarity of their work and making it suitable for eLife.

Abstract: "first time" Several times the authors mention this paper that this is the first time that such a study is done. It's only partially true. Over the past two years a lot of study have been published or are currently in accepted state about analysis of COVID-19 cardiac samples (for instance Walsh 2021 that the authors cite analyze partially whole heart as well).

l69: "we have introduce". The authors are not the only group working on this topic. A more general sentence saying that this field is growing would be more appropriate.

l76: "entire organ" -> "entire human organ". The example cited of Walsh is on human organs. However, a lot of work has been already done in the past in entire organ of animals such as mouse.

l79: "cytoarchitecture". It's a bit misleading as we expect to see results of the level of TEM. It's true that this level is reached for the results presented with the WG configuration but partially for the core analysis of the paper. Indeed the results focus on the general organization and vasculature.

Why the analysis has not been done in parallel or at the same time?

Is there any risk for this technique on future analysis due to sample degradation for instance?

l92: "based on visual impression" -> Sounds not scientific. Have the data been analyzed by a pathologist for assessment? Is that how it is done by pathologist?

l93: "automated image processing". Partially true, one still need to spend some times for doing manual segmentation. In the document nothing is mentioned about the quantity of images necessary for this process neither.

Figure 1: Later in the paper, the authors are describing the 3 methods used for the analysis as "LJ setup", "PB configuration", "WG configuration". First for clarity I would choose either setup either configuration. Then, I would also introduce those names in the figure or at least in the legend.

(B) -> "from one of the control and one of the Covid-19 samples". Otherwise it's misleading and difficult to understand that only one of each has been analyzed. How the choice actually has been made between the samples?

(D) plane of 3x3. Why 3x3 and not larger or smaller? Would a 360 or half scans be able to cover the interesting part of the biopsy that is in the end analyzes ? (i.e. to avoid the cropping due to the holder)?

(E) "was taken from a control sample". The analysis have been done on a COVID-19 sample as well, no?

Appendix Figure 1: The figure of the Haematoxylin and Eosin staining presents images too pink and no details of the microstructures can be really seen. Are those slices corresponding exactly to the samples before the biopsy punch? Have they been compared to slices for all samples on the X-ray datasets?

l113: "Biopsy punches". How the areas have been selected? Why the amount of biopsy punch are not identical for all samples? i.e. why not taking the entire block or selecting 2 samples per patients as for the CTRL?

l117: "one (Ctr) biopsy". One control and one COVID-19 are presented.

l123: It would be nice to have the source size of the liquid jet as well as the corresponding magnification factor. Where the broad spectrum used or only 9.25 keV? What is the true resolution compare to the pixel size?

l132: The measurement have been done at a different energy. What is the impact of such changes on the results? Would a higher energy be interesting?

l136: "the continuous scan mode". This is the property of the rotation stage I guess, and not the reason of being able to perform 3x3 tomography scans. Isn't it more a question of speed and stability of the stages? Have the scan been perform over 180 degrees?

l137: "dark field images were taken" I guess that flat field as well?

l138: "150 tomographic scans were recorded". It's not clear that this is for the entire amount of samples (with 3x3 for each).

l141: "1mm diameter biopsy punch". How this has been performed? Seems very tricky to extract 1mm rod from the 3.5mm one. How the area has been chosen. Was the height of the sample the same?

l139-153. It was not clear for me from the beginning that the 2 samples have been analyzed in 2 different configuration of the setup. A sentence introducing and explaining this would be helpful.

Table2: the source sample distance is missing for the PB configuration. Why the number of projections in the case of the PB configuration is the double than the 2 other configuration. Why the amount of empties/dark field so large? Empties is not the common name used in the field. Usually the term "flats" is preferred.

What means the acquisition time 3x0.6?

It would be interesting to give the total scan time for each technique.

l156: "local median filtering"

As the authors are using different phase contrast techniques for their analysis, a short introduction and a clearer organization of this paragraph would help in the comprehension for non-specialist. Some references could also be added. It is not easy to follow the difference between the reconstruction and phase retrieval techniques used for each technique.

l177: "datasets were binned by a factor of 2".

Tomographic datasets can indeed become very heavy, specially after stitching. However, here the purpose of using synchrotron is to reach higher resolution and higher throughput than in laboratory. When binning the data, the pixel size / resolution is also reduced. Therefore leading to a similar pixel size than in laboratory. Have the analysis be computed on the binned data or on the full datasets?

Table3: 1/ The d/b ratio values are real values of δ/β or just the result of the ratio? 2/ One can not understand what are the parameters.

l183: "corrupted datasets" What was the issues here? Need clarifications.

l197: "32 pixels for PB datasets" If the analysis has been done on binned data, it would lead to different size compare to the 12 pixels for the LJ acquisitions?

"A smoothing parameter of 2 pixels" It would be nice to have more explanation here. Have this been applied to both PB and LJ datasets?

l213: "the paraffin surrounding" Why the mask has been applied only on the LJ datasets. Isn't it included as well in the acquisition of the PB datasets?

l214: "Since one axis of the shape measure is redundant" Could the author clarify this point?

l231: "A small number of axis-aligned 2D slices was annotated". It would be very interesting to know the amount of annotated datasets that have been necessary to perform the analysis. Indeed, this is the most critical point when performing machine learning technique. Sparsely annotated data sets is a very promising technique, specially for analysing large amount of images. What was the percentage of images of annotated volumes were kept for training / validation?

l236: Why 96x96x96 voxels? What was the total size of the datasets?

l239 "256 subvolumes" How is this amount representative of the entire datasets?

l243: "A separate model" why was it necessary? Doesn't it create bias in the results?

l247: "adding additional annotations" Do I understand correctly that the analysis is run a first time, then the authors look at the images and visually correct some of them and re-run the analysis?

l314: "different areas of the same heart" How those areas have been chosen. Why 3 samples per patients?

l323: "dark stripes" In the figure it's easy to distinguish white stripes, but not the dark ones. Maybe small areas or markers would help to follow the description.

l329 / Figure 3: How to make the difference between paraffin cracks and inclusion compare to other features?

l337: "samples near an artery". How the areas have been chosen and what is the impact on the results presented?

l360: Appendix 1 Figure 2: would be nice to have arrows in the figure to understand what is erythrocytes and capillaries.

l368: "340x340x340um3": Why this size has been chosen and not the entire volume?

l371: "cytosol" would be nice to have this indicated in the figure.

l375: "nucleus" It seems that we can see only on in the figure. Do we see others in the 3D volume?

l386: "speeding-up the measurement sequence" It would be interesting to indeed grasp the difficulties of such measurement and time necessary.

Table 4: mean shape mu_p and mu_s are of the same order (ctrl or diseased) for mu_l there is a difference between control versus diseased however the σ is pretty large (almost 30% in certain case). Similar remarks for the elliptical fit (54% for σ in the case of Control…). The Authors are aware and found an interesting way of representing the results (Figure 5A). However, can the authors comments on why the errors are so huge, and how can one conclude in term of differences according to those results? Because even on the graph, we see tendency but to make strong concluding statement is difficult. I would therefore modify part of the eluding sentence in this direction.

Maybe could they compare with histology analysis conclude beforehand to help in the understanding.

l403: Authors are aware that their methods have some limitation. For instance "depend on tissue preservation and preparations".

l419: Welch t-test: Would be good to have a reference an explanation sentence.

l427: "Surface rendering" from which acquisition? Would be nice also to refer to the corresponding method paragraph.

l438: Could the authors explains a bit more what is the probability density function and how it is obtained. Figure 6E: The authors state clearly that their is a higher amount of branching points in the Covid19 sample. However, this is the results on one sample and the figure is not clearly stating that. For instance for 5 vertex degree the control seems superior than the covid.

l449: "first report" What about Walsch 2021?

l465: "non-destructively". It's not totally non-destructive technique as one still need to make a biopsy punch on the blocks.

l470: "conventional histopathology assessment" It would be nice to have a reference on the histology conventional analysis are adding in supplementary material correlation between X-ray images and histology images.

Figure 5: how those 2 samples have been chosen compare to the previous selection.

l493: "volume throughput of 10ˆ7 um3 /s. Maybe a simpler presentation would be more meaningful for the general audience, like acquisition time for one sample with which setup.

l497: rather than by photon flux: This is valid for the synchrotron acquisition I guess. But then what would be the purpose of the new synchrotron source if the flux is already not fully exploited? Same question concerning the following sentence about the attenuators used to prevent detector saturation. What about dose on the sample then?

l500: why so huge acquisition time range 200ms to 2500ms?

Appendix 1 Figure 3: In all the images we see like a cross in the images creating a white blur. Can the authors comment on that? Specially because grey levels are used for the analysis. How this effect can affect the results? A comment should also be added for the square missing (i.e. areas for corrupted files). Why measurements have not be redone locally?

Reviewer #2 (Recommendations for the authors):

It would be important to improve the samples' statistics.

https://doi.org/10.7554/eLife.71359.sa1

Author response

Reviewer #1 (Recommendations for the authors):

Below the authors will find my additional comments to increase the clarity of their work and making it suitable for eLife.

Abstract: "first time" Several times the authors mention this paper that this is the first time that such a study is done. It's only partially true. Over the past two years a lot of study have been published or are currently in accepted state about analysis of COVID-19 cardiac samples (for instance Walsh 2021 that the authors cite analyze partially whole heart as well).

Indeed there are multiple studies on structural changes of cardiac tissue due to Covid-19. However, this manuscript presents the first three-dimensional structural comparison of cardiac tissue from Covid-19 patients to different diseases. To avoid confusions we removed “first time” from the abstract.

l69: "we have introduce". The authors are not the only group working on this topic. A more general sentence saying that this field is growing would be more appropriate.

We have rephrased the sentence accordingly.

l76: "entire organ" -> "entire human organ". The example cited of Walsh is on human organs. However, a lot of work has been already done in the past in entire organ of animals such as mouse.

We have rephrased the sentence.

l79: "cytoarchitecture". It's a bit misleading as we expect to see results of the level of TEM. It's true that this level is reached for the results presented with the WG configuration but partially for the core analysis of the paper. Indeed the results focus on the general organization and vasculature.

We have rephrased “cytoarchitecture” to avoid confusions.

Why the analysis has not been done in parallel or at the same time?

Is there any risk for this technique on future analysis due to sample degradation for instance?

All samples have been investigated during the clinical routine and were provided for X-ray analysis. Formalin fixed paraffin embedded tissue are stable and do not degrade over time. At the relatively low dose invested, we have no indications of any sample structural degradation during or by the X-ray scans. However, we first wanted a clinical more established assessment before the X-ray analysis.

l92: "based on visual impression" -> Sounds not scientific. Have the data been analyzed by a pathologist for assessment? Is that how it is done by pathologist?

The data as well as the histology shown in the appendix was analyzed by pathologists. We have rephrased the sentence.

l93: "automated image processing". Partially true, one still need to spend some times for doing manual segmentation. In the document nothing is mentioned about the quantity of images necessary for this process neither.

The analysis of the laboratory data was performed automatically as described in the “Structure tensor analysis” section. The V-net based segmentation required manual annotation which acts as ground truth for the training of the network, but also here the statistical processing of the results based on graph theory was performed automated.

Figure 1: Later in the paper, the authors are describing the 3 methods used for the analysis as "LJ setup", "PB configuration", "WG configuration". First for clarity I would choose either setup either configuration. Then, I would also introduce those names in the figure or at least in the legend.

The wording has been changed to “setup” throughout, and is mentioned in the legends.

(B) -> "from one of the control and one of the Covid-19 samples". Otherwise it's misleading and difficult to understand that only one of each has been analyzed. How the choice actually has been made between the samples?

(D) plane of 3x3. Why 3x3 and not larger or smaller? Would a 360 or half scans be able to cover the interesting part of the biopsy that is in the end analyzes ? (i.e. to avoid the cropping due to the holder)?

(E) "was taken from a control sample". The analysis have been done on a COVID-19 sample as well, no?

We have now rephrased the caption to clarify the experimental setup and process of sample preparation and data acquisition.

Appendix Figure 1: The figure of the Haematoxylin and Eosin staining presents images too pink and no details of the microstructures can be really seen. Are those slices corresponding exactly to the samples before the biopsy punch? Have they been compared to slices for all samples on the X-ray datasets?

The micrographs shown in Appendix Figure 1 were recorded from all samples before a biopsy punch was taken, thus they correspond to the upper slice of the tomographic reconstructions. The presented H and E images represent the complete FFPE-sample. Both, histology and phase-contrast tomography data, were investigated by pathologists to describe the cardiac structure.

l113: "Biopsy punches". How the areas have been selected? Why the amount of biopsy punch are not identical for all samples? i.e. why not taking the entire block or selecting 2 samples per patients as for the CTRL?

The 3.5mm punches were selected to contain as much tissue as possible. We took one biopsy from each pathological tissue block which could be provided for this study. Fortunately, multiple control tissue blocks from the same donor could be provided, thus we were able to take multiple biopsies from the same patient.

l117: "one (Ctr) biopsy". One control and one COVID-19 are presented.

Indeed! We included the missing information in the text.

l123: It would be nice to have the source size of the liquid jet as well as the corresponding magnification factor. Where the broad spectrum used or only 9.25 keV? What is the true resolution compare to the pixel size?

The definition of the geometric magnification is provided in the text. We now also have included the corresponding magnifications in Table 2. For the phase-retrieval we have used the value of the Kα line. The resolution was not determined for this experiment, but we could show that the resolution is in the range of less than 2 px for this configuration in a different work. We now have included the respective reference (Reichardt et al., JMI, 2020).

l132: The measurement have been done at a different energy. What is the impact of such changes on the results? Would a higher energy be interesting?

In general, the interaction of X-ray radiation with matter is lower at higher energies. Moreover, the photon energy has to be optimized also with respect to the X-ray optics as undulator, monochromator, focusing optics and waveguide channel. Hence, different setups enable different photon energies, and optimizing these individually (for each setup) results in higher image quality.

l136: "the continuous scan mode". This is the property of the rotation stage I guess, and not the reason of being able to perform 3x3 tomography scans. Isn't it more a question of speed and stability of the stages? Have the scan been perform over 180 degrees?

That is correct. We have rephrased the paragraph and included the information on the angular range (360°!).

l137: "dark field images were taken" I guess that flat field as well?

Yes. The information was added.

l138: "150 tomographic scans were recorded". It's not clear that this is for the entire amount of samples (with 3x3 for each).

We acquired 3x3 datasets for 17 samples (153 tomograms), as we now explain in more detail.

l141: "1mm diameter biopsy punch". How this has been performed? Seems very tricky to extract 1mm rod from the 3.5mm one. How the area has been chosen. Was the height of the sample the same?

Indeed it takes a firm hand to take a 1mm biopsy from the 3.5 mm core. Thus, the small punch was extracted from the center of the 3.5mm biopsy punch. The height of both biopsies is the same.

l139-153. It was not clear for me from the beginning that the 2 samples have been analyzed in 2 different configuration of the setup. A sentence introducing and explaining this would be helpful.

The process of sample preparation and data acquisition is described in line 113-119

Table2: the source sample distance is missing for the PB configuration.

The source to sample distance was 88 m. We added the information to the table.

Why the number of projections in the case of the PB configuration is the double than the 2 other configuration. Why the amount of empties/dark field so large?

Since the data acquisition in the PB configuration can be performed in less than 2 minutes we extended the number of projections and flat field images.

Empties is not the common name used in the field. Usually the term "flats" is preferred.

We changes the name in the entire manuscript to flat field images.

What means the acquisition time 3x0.6?

Since the detector chip saturates at longer exposure times, we averaged 3 projections with an acquisition time of 0.6s. We now included a detailed description of the acquisition process.

It would be interesting to give the total scan time for each technique.

We have included the time for the different configurations.

l156: "local median filtering"

As the authors are using different phase contrast techniques for their analysis, a short introduction and a clearer organization of this paragraph would help in the comprehension for non-specialist. Some references could also be added. It is not easy to follow the difference between the reconstruction and phase retrieval techniques used for each technique.

We have restructured the paragraph, added some general information and included the reference for the toolbox used for the analysis at the beginning of the section.

l177: "datasets were binned by a factor of 2".

Tomographic datasets can indeed become very heavy, specially after stitching. However, here the purpose of using synchrotron is to reach higher resolution and higher throughput than in laboratory. When binning the data, the pixel size / resolution is also reduced. Therefore leading to a similar pixel size than in laboratory. Have the analysis be computed on the binned data or on the full datasets?

The analysis has been performed on the binned data. The difference in data quality is shown in Appendix Figure 2.

Table3: 1/ The d/b ratio values are real values of δ/β or just the result of the ratio? 2/ One can not understand what are the parameters.

The δ/β-ratio given in Tab.3 is the ratio which was used for phase retrieval (see Cloetens et al., 1999). In other words, they have to be regarded as effective parameters (guided by known or estimated values of the materials). All phase retrieval parameters given in this table are necessary and sufficient for replication starting from the raw data, along with the procedures explained in the corresponding references. A detailed description of the different phase retrieval algorithms would exceed the framework of this manuscript.

l183: "corrupted datasets" What was the issues here? Need clarifications.

Some scans were affected by errors in the data transfer, which was noticed only after the available beamtime.

l197: "32 pixels for PB datasets" If the analysis has been done on binned data, it would lead to different size compare to the 12 pixels for the LJ acquisitions?

"A smoothing parameter of 2 pixels" It would be nice to have more explanation here. Have this been applied to both PB and LJ datasets?

32 pixels refer to the unbinned data, as mentioned in the following sentence in the manuscript.

l213: "the paraffin surrounding" Why the mask has been applied only on the LJ datasets. Isn't it included as well in the acquisition of the PB datasets?

The statistical analysis based on the structure tensor has only be applied to the LJ data.

l214: "Since one axis of the shape measure is redundant" Could the author clarify this point?

The three shape measures add up to one, as described in line 210

l231: "A small number of axis-aligned 2D slices was annotated". It would be very interesting to know the amount of annotated datasets that have been necessary to perform the analysis. Indeed, this is the most critical point when performing machine learning technique. Sparsely annotated data sets is a very promising technique, specially for analysing large amount of images. What was the percentage of images of annotated volumes were kept for training / validation?

We did not keep exact account of the volume percentage where manual annotation was performed. In fact, manual segmentation of a single blood vessel can already suffice if the data quality is high and comparable between datasets.

l236: Why 96x96x96 voxels? What was the total size of the datasets?

Training and augmentation of data requires selection of small subvolumes, for two reasons: (1.) The net learns that the large configuration and borders are not important, and (2.) On the order of 10 subvolumes can be treated efficiently in parallel on the graphics card.

l239 "256 subvolumes" How is this amount representative of the entire datasets?

The subvolumes are randomly chosen. The segmentation results were validated by visual inspection.

l243: "A separate model" why was it necessary? Doesn't it create bias in the results?

For the analysis of the Covid-19 sample, a separate model needed to be trained, since the structure of the vasculature differs a lot from the controls: as described in the Results section, the blood vessels of the Covid-19 patients are mostly filled with blood while control vessels were mostly abundant of blood. In order to improve the quality of the segmentation the network had to be retrained for said differences.

l247: "adding additional annotations" Do I understand correctly that the analysis is run a first time, then the authors look at the images and visually correct some of them and re-run the analysis?

Yes, correct. This procedure improves the quality of the segmentation.

l314: "different areas of the same heart" How those areas have been chosen. Why 3 samples per patients?

Areas were chosen by eye and comparison to neighboring histological sections. 3 tissue blocks per patient (control group) were provided.

l323: "dark stripes" In the figure it's easy to distinguish white stripes, but not the dark ones. Maybe small areas or markers would help to follow the description.

The dark stripes (collagen sheets) are marked in Figure 3.

l329 / Figure 3: How to make the difference between paraffin cracks and inclusion compare to other features?

This is distinguished based on gray values (electron density).

l337: "samples near an artery". How the areas have been chosen and what is the impact on the results presented?

Samples were chosen based on the histology, in order to increase tissue content in the punch. In some cases it simply turned out that a medium size vessel was present in the chosen volume.

l360: Appendix 1 Figure 2: would be nice to have arrows in the figure to understand what is erythrocytes and capillaries.

This figure is intended only to compare the data quality between synchrotron and in-house; we would prefer not to distract from this by any additional labels.

l368: "340x340x340um3": Why this size has been chosen and not the entire volume?

The FOV is limited by the effective pixelsize and the amount of detector pixels (which is limited by its manufacturing).

l371: "cytosol" would be nice to have this indicated in the figure.

The myofibrils are located in the cytosol (background); we prefer to only indicate the myofibril here.

l375: "nucleus" It seems that we can see only on in the figure. Do we see others in the 3D volume?

Each cardiomyocyte within the volume contains one nucleus. In the figure we present one exemplary region, the high resolution data is publicly available and can be further investigated. Indeed, many more nuclei can be identified!

l386: "speeding-up the measurement sequence" It would be interesting to indeed grasp the difficulties of such measurement and time necessary.

For future work we plan to update the instrument (waveguide optics, automated alignment, continuous data acquisition).

Table 4: mean shape mu_p and mu_s are of the same order (ctrl or diseased) for mu_l there is a difference between control versus diseased however the σ is pretty large (almost 30% in certain case). Similar remarks for the elliptical fit (54% for σ in the case of Control…). The Authors are aware and found an interesting way of representing the results (Figure 5A). However, can the authors comments on why the errors are so huge, and how can one conclude in term of differences according to those results? Because even on the graph, we see tendency but to make strong concluding statement is difficult. I would therefore modify part of the eluding sentence in this direction.

Maybe could they compare with histology analysis conclude beforehand to help in the understanding.

The large intra-group variance reflects the pronounced variability between individual subjects and samples, which is in line with experience of conventional histology. We have added a comment on this.

l403: Authors are aware that their methods have some limitation. For instance "depend on tissue preservation and preparations".

Sample preparation is always an issue when investigating biological tissues and it is important to take the sample quality into account.

l419: Welch t-test: Would be good to have a reference an explanation sentence.

Since the Welch t-test is a concept which was first published in 1947 and used as gold standard for statistical investigations we refrained from adding a reference.

l427: "Surface rendering" from which acquisition? Would be nice also to refer to the corresponding method paragraph.

We included the information on the corresponding datasets in the text and on the software used for visualization (avizo) in the methods section.

l438: Could the authors explains a bit more what is the probability density function and how it is obtained. Figure 6E: The authors state clearly that their is a higher amount of branching points in the Covid19 sample. However, this is the results on one sample and the figure is not clearly stating that. For instance for 5 vertex degree the control seems superior than the covid.

The PDF quantifies the node connectivity after skeletonization procedure, which is carried out according to ref. [Lee et al., 1994]. The PDF is normalized, all red (or green) columns add up to one. Now the different heights are a ‘fingerprint’ of the graph topology. The 5 vertex degree is negligible (extremely rare), and even 4 is marginal. This mainly serves to show that vertex degrees 1-3 are relevant here. This is now explained more clearly.

l449: "first report" What about Walsch 2021?

We present the first report comparing cardiac histopathology of Covid-19 to influenza and coxsackie myocarditis; we now state more clearly that Walsh et al., have already studied heart tissue in Covid-19, albeit with different sample preparation and at different resolution.

l465: "non-destructively". It's not totally non-destructive technique as one still need to make a biopsy punch on the blocks.

We removed non-destructively in this context. In general X-ray tomography is a destruction-free technique and the punches can be placed back into the blocks and/or be used for further investigations.

l470: "conventional histopathology assessment" It would be nice to have a reference on the histology conventional analysis are adding in supplementary material correlation between X-ray images and histology images.

We have made this important point in our previous manuscript on lung damage in Covid-19, and include this reference now in the methods section.

Figure 5: how those 2 samples have been chosen compare to the previous selection.

The samples in Figure 6 were chosen randomly.

l493: "volume throughput of 10ˆ7 um3 /s. Maybe a simpler presentation would be more meaningful for the general audience, like acquisition time for one sample with which setup.

Acquisition time per sample is now also given in the manuscript.

l497: rather than by photon flux: This is valid for the synchrotron acquisition I guess. But then what would be the purpose of the new synchrotron source if the flux is already not fully exploited? Same question concerning the following sentence about the attenuators used to prevent detector saturation. What about dose on the sample then?

This is true for PB acquisition, but not the WG configuration. This is now stated more precisely.

l500: why so huge acquisition time range 200ms to 2500ms?

The photon flux depends on the diameter of the waveguide channel, thus the acquisition time has to be adapted to achieve similar fluence.

Appendix 1 Figure 3: In all the images we see like a cross in the images creating a white blur. Can the authors comment on that? Specially because grey levels are used for the analysis. How this effect can affect the results? A comment should also be added for the square missing (i.e. areas for corrupted files). Why measurements have not be redone locally?

The intensity changes result from local tomography artefacts and appear in the stitched datasets. Since the structural analysis works not on the gray value but on its gradients, the effect can be neglected in the shape measure analysis. The information of corrupted datasets was included in the caption.

Reviewer #2 (Recommendations for the authors):

It would be important to improve the samples' statistics.

In this work, we investigated 26 cardiac tissue samples from various diseases (Covid-19, coxsackie myocarditis and H1N1/A influenza) at the laboratory setup. More than 500GB raw data and for each volume approximately 2000 voxels cubed were analyzed in terms of structural shape measure characterization. In follow up studies, we plan to extend the sample statistics and use extensive mathematical models based on optimal transport to refine structural differences between individual sample groups. However, this relies on the availability of autopsy samples, which are always quite limited, since autopsies are not routinely performed of patients who succumbed to Covid-19.

https://doi.org/10.7554/eLife.71359.sa2

Article and author information

Author details

  1. Marius Reichardt

    Institut für Röntgenphysik, Georg-August-Universität Göttingen, Friedrich-Hund-Platz, Göttingen, Germany
    Contribution
    Conceptualization, Investigation, Methodology, Software, Visualization, Writing - original draft, Writing – review and editing
    Competing interests
    No competing interests declared
  2. Patrick Moller Jensen

    Technical University of Denmark, Richard Petersens Plads, Kopenhagen, Denmark
    Contribution
    Investigation, Methodology, Resources, Software, Visualization, Writing - original draft, Writing – review and editing
    Competing interests
    No competing interests declared
  3. Vedrana Andersen Dahl

    Technical University of Denmark, Richard Petersens Plads, Kopenhagen, Denmark
    Contribution
    Methodology, Resources, Supervision, Writing – review and editing
    Competing interests
    No competing interests declared
  4. Anders Bjorholm Dahl

    Technical University of Denmark, Richard Petersens Plads, Kopenhagen, Denmark
    Contribution
    Methodology, Resources, Supervision, Writing – review and editing
    Competing interests
    No competing interests declared
  5. Maximilian Ackermann

    Institute of Anatomy and Cell Biology, University Medical Center of the Johannes Gutenberg-University Mainz, Mainz, Germany
    Contribution
    Investigation, Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9996-2477
  6. Harshit Shah

    1. Medizinische Hochschule Hannover (MHH), Hannover, Germany
    2. Deutsches Zentrum für Lungenforschung (DZL), Hannover (BREATH), Hannover, Germany
    Contribution
    Investigation
    Competing interests
    No competing interests declared
  7. Florian Länger

    1. Medizinische Hochschule Hannover (MHH), Hannover, Germany
    2. Deutsches Zentrum für Lungenforschung (DZL), Hannover (BREATH), Hannover, Germany
    Contribution
    Investigation, Resources
    Competing interests
    No competing interests declared
  8. Christopher Werlein

    Medizinische Hochschule Hannover (MHH), Hannover, Germany
    Contribution
    Investigation, Methodology, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7694-4257
  9. Mark P Kuehnel

    1. Medizinische Hochschule Hannover (MHH), Hannover, Germany
    2. Deutsches Zentrum für Lungenforschung (DZL), Hannover (BREATH), Hannover, Germany
    Contribution
    Investigation, Methodology, Resources, Writing – review and editing
    Competing interests
    No competing interests declared
  10. Danny Jonigk

    1. Medizinische Hochschule Hannover (MHH), Hannover, Germany
    2. Deutsches Zentrum für Lungenforschung (DZL), Hannover (BREATH), Hannover, Germany
    Contribution
    Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing – review and editing
    For correspondence
    Jonigk.Danny@mh-hannover.de
    Competing interests
    No competing interests declared
  11. Tim Salditt

    Institut für Röntgenphysik, Georg-August-Universität Göttingen, Friedrich-Hund-Platz, Göttingen, Germany
    Contribution
    Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing - original draft, Writing – review and editing
    For correspondence
    tsaldit@gwdg.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4636-0813

Funding

Bundesministerium für Bildung und Forschung (Max Planck School Matter to Life)

  • Marius Reichardt
  • Tim Salditt

Bundesministerium für Bildung und Forschung (05K19MG2)

  • Tim Salditt

Deutsche Forschungsgemeinschaft (EXC 2067/1-390729940)

  • Tim Salditt

H2020 European Research Council (XHale)

  • Danny Jonigk

Deutsche Forschungsgemeinschaft (KFO311 (project Z2))

  • Danny Jonigk

Hanseatic League of Science

  • Patrick Moller Jensen

H2020 European Research Council (771883)

  • Danny Jonigk

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Ove Hansen for help with deep learning, Markus Osterhoff, Michel Sprung, and Fabian Westermeier for support at P10. Florian Länger for helpful discussion, Patrick Zardo for providing control specimen, and Bastian Hartmann, Jan Goemann, Regina Engelhardt, Anette Müller-Brechlin and Christina Petzold for their excellent technical help. It is also acknowledge DESY photon science management for the Covid-19 beamtime call and the granted beamtime. Funding This research was supported by the Max Planck School Matter to Life supported by the German Federal Ministry of Education and Research (BMBF) in collaboration with the Max Planck Society (MR,TS), as well as BMBF grant No. 05K19MG2 (TS), German Research Foundation (DFG) under Germanys Excellence Strategy -EXC 2067/1–390729940 (TS), the European Research Council Consolidator Grant XHale, 771,883 (DJ) and KFO311 (project Z2) of the DFG (DJ). Participation of PMJ was supported by a HALOS exchange stipend.

Ethics

Human subjects: Formalin-fixed paraffin-embedded tissue blocks of control hearts, influenza and coxsackie virus myocarditis hearts were retrieved from archived material from the Institute of Pathology at Hannover Medical School in accordance with the local ethics committee (ethics vote number: 1741-2013 and 2893-2015). Formalin-fixed paraffin-embedded tissue blocks of COVID-19 autopsy cases were retrieved after written consent in accordance with the local ethics committee at Hannover medical school (ethics vote number: 9022 BO K 2020).

Senior Editor

  1. Matthias Barton, University of Zurich, Switzerland

Reviewing Editor

  1. Hina Chaudhry, Harvard University, United States

Publication history

  1. Received: June 17, 2021
  2. Preprint posted: September 18, 2021 (view preprint)
  3. Accepted: December 10, 2021
  4. Accepted Manuscript published: December 21, 2021 (version 1)
  5. Version of Record published: January 10, 2022 (version 2)

Copyright

© 2021, Reichardt et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 4,379
    Page views
  • 527
    Downloads
  • 4
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Marius Reichardt
  2. Patrick Moller Jensen
  3. Vedrana Andersen Dahl
  4. Anders Bjorholm Dahl
  5. Maximilian Ackermann
  6. Harshit Shah
  7. Florian Länger
  8. Christopher Werlein
  9. Mark P Kuehnel
  10. Danny Jonigk
  11. Tim Salditt
(2021)
3D virtual histopathology of cardiac tissue from Covid-19 patients based on phase-contrast X-ray tomography
eLife 10:e71359.
https://doi.org/10.7554/eLife.71359
  1. Further reading

Further reading

    1. Epidemiology and Global Health
    2. Medicine
    Qing Shen, Huan Song ... Unnur Valdimarsdóttir
    Research Article Updated

    Background:

    The association between cardiovascular disease (CVD) and selected psychiatric disorders has frequently been suggested while the potential role of familial factors and comorbidities in such association has rarely been investigated.

    Methods:

    We identified 869,056 patients newly diagnosed with CVD from 1987 to 2016 in Sweden with no history of psychiatric disorders, and 910,178 full siblings of these patients as well as 10 individually age- and sex-matched unrelated population controls (N = 8,690,560). Adjusting for multiple comorbid conditions, we used flexible parametric models and Cox models to estimate the association of CVD with risk of all subsequent psychiatric disorders, comparing rates of first incident psychiatric disorder among CVD patients with rates among unaffected full siblings and population controls.

    Results:

    The median age at diagnosis was 60 years for patients with CVD and 59.2% were male. During up to 30 years of follow-up, the crude incidence rates of psychiatric disorder were 7.1, 4.6, and 4.0 per 1000 person-years for patients with CVD, their siblings and population controls. In the sibling comparison, we observed an increased risk of psychiatric disorder during the first year after CVD diagnosis (hazard ratio [HR], 2.74; 95% confidence interval [CI], 2.62–2.87) and thereafter (1.45; 95% CI, 1.42–1.48). Increased risks were observed for all types of psychiatric disorders and among all diagnoses of CVD. We observed similar associations in the population comparison. CVD patients who developed a comorbid psychiatric disorder during the first year after diagnosis were at elevated risk of subsequent CVD death compared to patients without such comorbidity (HR, 1.55; 95% CI, 1.44–1.67).

    Conclusions:

    Patients diagnosed with CVD are at an elevated risk for subsequent psychiatric disorders independent of shared familial factors and comorbid conditions. Comorbid psychiatric disorders in patients with CVD are associated with higher risk of cardiovascular mortality suggesting that surveillance and treatment of psychiatric comorbidities should be considered as an integral part of clinical management of newly diagnosed CVD patients.

    Funding:

    This work was supported by the EU Horizon 2020 Research and Innovation Action Grant (CoMorMent, grant no. 847776 to UV, PFS, and FF), Grant of Excellence, Icelandic Research Fund (grant no. 163362-051 to UV), ERC Consolidator Grant (StressGene, grant no. 726413 to UV), Swedish Research Council (grant no. D0886501 to PFS), and US NIMH R01 MH123724 (to PFS).

    1. Epidemiology and Global Health
    Bingyi Yang, Bernardo García-Carreras ... Derek A Cummings
    Research Article

    Background: Over a life-course, human adaptive immunity to antigenically mutable pathogens exhibits competitive and facilitative interactions. We hypothesize that such interactions may lead to cyclic dynamics in immune responses over a lifetime.

    Methods: To investigate the cyclic behavior, we analyzed hemagglutination inhibition titers against 21 historical influenza A(H3N2) strains spanning 47 years from a cohort in Guangzhou, China and applied Fourier spectrum analysis. To investigate possible biological mechanisms, we simulated individual antibody profiles encompassing known feedbacks and interactions due to generally recognized immunological mechanisms.

    Results: We demonstrated a long-term periodicity (about 24 years) in individual antibody responses. The reported cycles were robust to analytic and sampling approaches. Simulations suggested that individual-level cross-reaction between antigenically similar strains likely explain the reported cycle. We showed that the reported cycles are predictable at both individual and birth-cohort level and that cohorts show a diversity of phases of these cycles. Phase of cycle was associated with the risk of seroconversion to circulating strains, after accounting for age and pre-existing titers of the circulating strains.

    Conclusions: Our findings reveal the existence of long-term periodicities in individual antibody responses to A(H3N2). We hypothesize that these cycles are driven by pre-existing antibody responses blunting responses to antigenically similar pathogens (by preventing infection and/or robust antibody responses upon infection), leading to reductions in antigen specific responses over time until individual's increasing risk leads to an infection with an antigenically distant enough virus to generate a robust immune response. These findings could help disentangle cohort-effects from individual-level exposure histories, improve our understanding of observed heterogeneous antibody responses to immunizations, and inform targeted vaccine strategy.

    Funding: This study was supported by grants from the NIH R56AG048075 (D.A.T.C., J.L.), NIH R01AI114703 (D.A.T.C., B.Y.), the Wellcome Trust 200861/Z/16/Z (S.R.) and 200187/Z/15/Z (S.R.). This work was also supported by research grants from Guangdong Government HZQB-KCZYZ-2021014 and 2019B121205009 (Y.G. and H.Z.). D.A.T.C., J.M.R. and S.R. acknowledge support from the National Institutes of Health Fogarty Institute (R01TW0008246). J.M.R. acknowledges support from the Medical Research Council (MR/S004793/1) and the Engineering and Physical Sciences Research Council (EP/N014499/1). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.