Deciphering Tumour Tissue Organization by 3D Electron Microscopy and machine learning

Abstract

Despite recent progress in the characterization of tumour components, the tri-dimensional (3D) organization of this pathological tissue and the parameters determining its internal architecture remain elusive. Here, we analysed the spatial organization of patient-derived xenograft tissues generated from hepatoblastoma, the most frequent childhood liver tumour, by serial block-face scanning electron microscopy using an integrated workflow combining 3D imaging, manual and machine learning-based semi-automatic segmentations, mathematics and infographics. By digitally reconstituting an entire hepatoblastoma sample with a blood capillary, a bile canaliculus-like structure, hundreds of tumour cells and their main organelles (e.g. cytoplasm, nucleus, mitochondria), we report unique 3D ultrastructural data about the organization of tumour tissue. We found that the size of hepatoblastoma cells correlates with the size of their nucleus, cytoplasm and mitochondrial mass. We also found anatomical connections between the blood capillary and the planar alignment and size of tumour cells in their 3D milieu. Finally, a set of tumour cells polarized in the direction of a hot spot corresponding to a bile canaliculus-like structure. In conclusion, this pilot study allowed the identification of bioarchitectural parameters that shape the internal and spatial organization of tumours, thus paving the way for future investigations in the emerging onconanotomy field.

Full text

ARTICLE
Deciphering tumour tissue organization by 3D
electron microscopy and machine learning
Baudouin Denis de Senneville1,2,11, Fatma Zohra Khoubai 3,11, Marc Bevilacqua4, Alexandre Labedade 5,
Kathleen Flosseau6, Christophe Chardot7, Sophie Branchereau8, Jean Ripoche9, Stefano Cairo 6,10,12,
Etienne Gontier 4,12 ✉& Christophe F. Grosset 3,12 ✉
Despite recent progress in the characterization of tumour components, the tri-dimensional
(3D) organization of this pathological tissue and the parameters determining its internal
architecture remain elusive. Here, we analysed the spatial organization of patient-derived
xenograft tissues generated from hepatoblastoma, the most frequent childhood liver tumour,
by serial block-face scanning electron microscopy using an integrated workflow combining
3D imaging, manual and machine learning-based semi-automatic segmentations, mathe-
matics and infographics. By digitally reconstituting an entire hepatoblastoma sample with a
blood capillary, a bile canaliculus-like structure, hundreds of tumour cells and their main
organelles (e.g. cytoplasm, nucleus, mitochondria), we report unique 3D ultrastructural data
about the organization of tumour tissue. We found that the size of hepatoblastoma cells
correlates with the size of their nucleus, cytoplasm and mitochondrial mass. We also found
anatomical connections between the blood capillary and the planar alignment and size of
tumour cells in their 3D milieu. Finally, a set of tumour cells polarized in the direction of a hot
spot corresponding to a bile canaliculus-like structure. In conclusion, this pilot study allowed
the identification of bioarchitectural parameters that shape the internal and spatial organi-
zation of tumours, thus paving the way for future investigations in the emerging
onconanotomy field.
https://doi.org/10.1038/s42003-021-02919-z OPEN
1CNRS, University of Bordeaux, “Institut de Mathématiques de Bordeaux”(IMB), UMR5251, 351 cours Libération, F-33400 Talence, France. 2INRIA Bordeaux,
MONC team, 200 av Vieille Tour, F-33400 Talence, France. 3Univ. Bordeaux, INSERM, Biotherapy of Genetic Diseases, Inflammatory Disorders and Cancer,
BMGIC, U1035, MIRCADE team, 146 rue Léo Saignat, 33076 Bordeaux, France. 4Univ. Bordeaux, CNRS, INSERM, Bordeaux Imaging Centre, BIC, UMS 3420,
US 4, F-33000 Bordeaux, France. 5Alexandre Labedade, freelance, F-33870 Vayres, France. 6XenTech Company, Genopole Campus 3, 4 Rue Pierre Fontaine,
F-91000 Évry-Courcouronnes, France. 7Chirurgie Pédiatrique-Transplantation, Hôpital Necker-Enfants Malades, Assistance Publique Hôpitaux de Paris, 149
rue de Sèvres, F-75015 Paris, France. 8Paediatric Surgery and Liver Transplant unit, Hôpital Bicêtre, Assistance Publique Hôpitaux de Paris –Université Paris
Saclay, 78 rue du Général Leclerc, F-94270 Le Kremlin-Bicêtre, France. 9Univ. Bordeaux, INSERM, retired, 146 rue Léo Saignat, 33076 Bordeaux, France.
10 Istituto di Ricerca Pediatrica, Corso Stati Uniti 4, 35127 Padova, Italy.
11
These authors contributed equally: Baudouin Denis de Senneville, Fatma Zohra
Khoubai.
12
These authors jointly supervised this work: Stefano Cairo, Etienne Gontier, Christophe F. Grosset. ✉email: etienne.gontier@u-bordeaux.fr;
christophe.grosset@inserm.fr
COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio 1
1234567890():,;
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Diagnostic, prognostic and predictive clinical cancer studies
are mainly based on the analysis of tissue sections in two
dimensions (2D). However, tumours often present a
complex architecture that 2D imaging cannot capture. Tumour cells
establish composite interactions with the surrounding cancerous cells,
extracellular matrix and stromal components, including blood
capillaries and immune cells. In recent investigations, we noted that
various hepatoma cells engrafted on the chick embryo chorioallantoic
membrane (CAM) were arranged histologically and microscopically
in a lineage-specificwaywhilegrowinginacomparablecontrolled
environment1,2. Moreover, pathological tumour atlases display a
multitude of 2D images with tumour tissues whose cells are orga-
nized differently while belonging to the same category of cancer
(http://www.pathologyatlas.ro;https://www.proteinatlas.org;https://
atlases.muni.cz). These observations suggest that the structural
organization of tumour tissue changes from one sample to another
andislikelycontrolledbyenvironmental and genetic factors such as
the tissue origin, state of differentiation, genetic programme, the
mutations they harbour, the activated molecular pathways, and
the surrounding stromal components such as blood capillaries.
Moreover, these interactions are dynamic and vary as the cancer
progresses and metastasizes. To shed light on this critical issue, we
investigated the ultrastructural pattern of cancer tissue.
The applications of 3D electron microscopy (EM) are still
under development. Since its advent in the 1930s, EM has allowed
the in-depth analysis of a wide range of biological samples.
Transmission electron microscopy (TEM) and scanning electron
microscopy (SEM) are routinely used to collect ultrastructural
data from biological specimens at nanoscale resolution and to
correlate structural images with biological functions. Although
effective, these techniques produce images only in two dimen-
sions. Therefore, investigating a large volume (100–1000 pL) of
biological tissue in 3D at a high resolution was hardly possible
until the development of 3D EM technologies including serial
block-face (SBF)-SEM and Focused Ion Beam-SEM3–6.
Hepatoblastoma (HB) is the most common form of liver cancer
in young children. Recently, we reported the classification of these
paediatric tumours in three groups, the use of Velcade®as a new
therapeutic option for the treatment of aggressive HB and the
development of an HB model in chick embryo for biological
studies1,7,8. We also developed HB-patient-derived xenografts
(PDX), a tumour model that recapitulates the histological, genetic
and biological characteristics of parental HB9. Here, we used SBF-
SEM to investigate the 3D internal organization of HB samples
obtained from HB PDXs. Following image acquisition by SBF-
SEM, we gathered quantitative data about the ultrastructure of
tumour tissue using 3D imaging tools, mathematical algorithms
and deep-learning approaches. We accurately measured the size
of tumour cells and their main subcellular components (e.g.
nucleus, mitochondria and cytoplasm), their planar alignment,
their spatial orientation and the distance between the cells and a
blood capillary. Results validate the relevance of our integrated
workflow using wide-field 3D EM and computational approaches
to investigate the deep internal architecture of HB tissue and the
spatial organization of the tumour cells. These structural and
functional data pave the way for future investigations in the
emerging field of onconanotomy.
Results
Image acquisition by EM and SBF-SEM of HB PDX tissues.To
study the 3D organization of HB-PDX tissues, we adapted the
protocol described by Deerinck et al.10 to tumour samples
(Supplementary Fig. 1) and analysed the ultrastructural mor-
phology and the region of interest (ROI) of each stained sample
by TEM using ultrathin sections. Figure 1a shows that the
staining of the four samples was uniform and well contrasted and
that all HB tissues retained their structure of origin. For example,
membranes, mitochondria, endoplasmic reticulum, nuclei, lipid
droplets and circulating red blood cells into capillaries (“bc”) were
clearly identifiable and typical with an excellent grey tone scale.
The morphology of endothelial and immune cells was typical.
Intercellular spaces were also visible and identified as bile
canaliculus-like structures (“bi”, sample 2)11,12, which are indi-
cative of liver-derived tissue. Overall, these results demonstrated
the preservation of the ultrastructure of our HB samples following
fixation and staining procedures.
Next, we analysed two HB PDXs by SBF-SEM and acquired
high-resolution images in Xand Yalong the Z-axis to produce a
substantial volumetric view of each sample. To obtain both image
quality and detailed volumetric information, the acquired volume
Fig. 1 Analysis of HB PDXs by EM. a Analysis of four HB PDX by TEM at
different magnifications. White dotted-line frame: Region-of-interest
corresponding to immediately higher magnification. bc, blood capillary; bi,
bile canaliculus-like structure; EC, endothelial cells; MA, macrophage; lv,
lipid vesicles; mb; plasma membrane; mi, mitochondrion; n, nucleus; ni,
nuclear invagination; nl, nucleolus; LY, lymphocyte; RC, red cells; TC,
tumour cell. bTop panels: SBF-SEM stacks from two samples. Bottom
panels: Extracted images of orthogonal plane from different faces of
reconstructed volume (XY,YZ,XZ), whose positions are displayed on
orthoslice view positions inside stack. Scale bar =10 µm.
ARTICLE COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z
2COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio
Content courtesy of Springer Nature, terms of use apply. Rights reserved

was about 70 µm × 70 µm × 25 µm (pixel time =10 µs, pixel size =
15 nm). After acquisition, the images were aligned to obtain 3D
stacks (Fig. 1b; Sample 1: 250 images, 113.9 pL; Sample 2: 246
images, 121.6 pL). While volumetric image resolution was not
isotropic (15 nm lateral resolution versus 100 nm depth resolution),
ultrastructural features within the 3D stack were perfectly
maintained and identifiable, allowing us to visualize the cells and
their subcellular components within the tumour samples from any
angleofviewonthethreeaxes(Fig.1b).
3D organization of individual tissue components. Following
stack reconstruction, we analysed the tissue features in tumour
Sample 1 (Supplementary Fig. 2a, b, Supplementary Video 1).
Using Ilastik13 and VAST-Lite14 software, we manually seg-
mented a blood capillary and its circulating red and white blood
cells (one leucocyte and one structure evoking a dendritic cell or a
platelet cluster) (Supplementary Fig. 2c, f, Supplementary
Video 1). The volume of the blood capillary portion irrigating the
tissue was 11.3 pL and the 11 entirely segmented murine red
blood cells were 61.92 ± 16.26 µm3in size (Supplementary
Fig. 2c–f, Supplementary Data 1 and Supplementary Video 1).
Next, we manually segmented a tumour cell from Sample 2 by
delimitating the cytoplasm, the nucleus and the mitochondria.
Again, the 3D digital representation allowed volumetric quanti-
fication of the cell and its cytoplasm, nucleus and mitochondrial
network (970.6, 657.1, 313.6 and 64.6 µm3, respectively), as well
as affording a morphologic view of the tumour cell in its 3D
environment (Supplementary Fig. 3; Supplementary Video 2).
The size of this tumour cell was consistent with the size range of
HB cells15,16.
Reconstruction of tumour 3D organization by semi-automated
segmentation. Although it produces accurate data, manual seg-
mentation is too laborious and time-consuming to allow the
analysis of hundreds of cells constituting a stack of about 110–120
pL. To overcome this problem and anticipate the analysis of
larger volumes of tumour (>200 pL), we implemented a semi-
automatic segmentation procedure comprising the manual seg-
mentation of the cytoplasm and nucleus of cells on 1 out of 10
images in the stack form Sample 2 and an automatic segmenta-
tion algorithm of ROI by proximal propagation along the Zaxis
(Supplementary Fig. 4). This procedure allowed us to segment
partially and entirely 182 tumour cells, 113 nuclei, 1 immune cell
and 3 portions of the same blood capillary (Fig. 2a–c, Supple-
mentary Fig. 5, Supplementary Video 3). The elongated shape of
the immune cell, the presence of many vacuoles in its cytoplasm
and its location near the blood capillary may suggest of a tumour-
infiltrating macrophage (Supplementary Fig. 5, Supplementary
Video 3). Following the alignment in a plane of 47 tumour cells
with a complete nucleus, we found that 36 tumour cells (76.6%)
and 35 nuclei (74.5%) had an inclination angle of 0 to 20°
(relative to the best alignment plane, see Online Methods), while
the inclination angle of the blood capillary was 8° (Fig. 2d, e,
Supplementary Fig. 6, Supplementary Data 1, Supplementary
Video 4). These data suggested that the blood capillary may
influence the spatial arrangement of tumour cells.
Next, we measured the tumour cell polarity (in the sense
of attraction toward a particular object) using a vectoral approach
(see Online Methods). In both cases, data showed that a set
of tumour cells polarized in the direction of a hot spot
corresponding to a bile canaliculus-like structure (Fig. 3,
Supplementary Fig. 7, Supplementary Video 5)11,12. Numerous
membrane protrusions were visible at the boundary between
the bile canaliculus-like structure and the canalicular membrane
of the surrounding tumour cells (Supplementary Fig. 7,
Supplementary Video 5). This structure into which hepatocytes
excrete their metabolic agents after enzymatic neutralization is
specific to normal liver tissue15 and has been reported to be
remnant in some HB11,12.
Finally, we focused our attention on the 21 fully segmented
tumour cells contained in sample 2. Mitochondria are key
organelles of tumour cell function but too numerous (between
200 and 400 per cell) and small in size to be manually segmented
from our whole tissue sample. Thus, a deep-learning algorithm
was fed using data from mitochondria manually segmented on
one single cell. Next, we performed two successive cycles of
automatic segmentation by deep-learning propagation combined
with manual segmentation clean-up of segmented mitochondria
(see Methods). Following this segmentation procedure, the size of
these cells and their cytoplasm, nucleus and mitochondrial
network was measured and component/cell ratios were calculated
(Supplementary Fig. 8a, b, Supplementary Data 1). In line with
the known relationship between the size or number of organelles
and the size of the cell17,18, we found that the larger the tumour
cell, the larger its cytoplasm, global mitochondrial network and
nucleus (Fig. 4a, Supplementary Data 1). By measuring the
distance between each of these cellular components and the blood
capillary, we also found that the tumour cells located near the
blood capillary were significantly larger than those located away
from it (Fig. 4b, c, Supplementary Data 1, Supplementary
Video 6). This inverse correlation with the distance to the
capillary was also observed when considering the cytoplasm and
the mitochondrial network (the latter having the highest
correlative value) but not the nuclei (Fig. 4c, Supplementary
Fig. 9a–c, Supplementary Data 1, Supplementary Video 6).
Altogether, these data suggested that the size of HB cells and their
mitochondrial content is linked to the distance with blood
capillary. These observations are in agreement with the decrease
in mitochondrial mass in cells exposed to low oxygen supply19,20.
Discussion
A still unanswered question is whether cells are distributed ran-
domly in cancer tissue or if and how their organization is gov-
erned by physical and molecular factors reminiscent of the
structure of normal tissue. In this pilot study, we used a 3D EM
approach named onconanotomy, i.e. a workflow to analyse in
depth the architecture of tumour samples using a combination of
sample preparation, SBF-SEM imaging, computational approa-
ches and infographic tools (see Technical workflow in Fig. 5). To
the best of our knowledge, this is the first report investigating the
complete ultrastructure of human tumour xenograft tissue by
high-resolution 3D EM.
While automated segmentation with machine learning meth-
ods has become an essential technique for analysing the ultra-
structure of all tissue and cell types, the main difficulty in our
work laid in the construction of the learning database. At the very
beginning of our study, we did not have any organelle segmen-
tation available and we used manual segmentation. Thus, we
developed semi-automatic approaches to accelerate the process.
One can anticipate that a fully automated segmentation of major
organelles (cytoplasm, nuclei, mitochondria, etc.), cell types
(blood cells, immune cells, hepatocytes and tumour cells) and
tissue structures (capillary, bile-canaliculus, etc.) is feasible (see
Segmented tissue elements in Fig. 5) based on sufficiently
populated training databases and given high potential of novel
technologies in artificial intelligence.
Previous detailed reports on the 3D bioarchitectural organi-
zation of tumour are lacking to fully establish if our data are
reflecting the real situation found in this pathological tissue.
However, we found red and white cells circulating in a blood
COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z ARTICLE
COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio 3
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Fig. 2 Alignment of HB cells and nuclei. a Sample 2 stack. b,cDigital representation of (a) with 182 cells (b) and 113 nuclei (c). dAlignment of cells and
blood capillary in a plane: (i) coloured sticks =main axis of tumour cells; red stick =main axis of blood capillary. (ii–ix) Angles between each main cell axis
and the best alignment plane. 2D cross-sectional maps for increasing depth along Z-axis. Scale bar =10 µm. b–dBlood capillary portions are in red.
eHistograms of cells (left panel) and nuclei (right panel) alignment angles. Red dashed line: alignment angle of blood capillary.
ARTICLE COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z
4COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Fig. 3 Tumour cell cluster with polarized shape orientation. Only tumour cells with a complete nucleus are considered here. Blood capillary portions are in
red. aAccumulation map of virtual rays [Accumulated ray intensity [a.u]; voxel-by-voxel basis; see yellowish voxels] emitted by each cell along its main
axis. (i) Reconstructed 3D image of accumulation map of virtual rays. (ii–ix) 2D cross-sectional maps for increasing depth along the Z-axis. bBinary
classification of polarised/unpolarised cells: (i) Reconstructed 3D image of polarised/unpolarised cells. 17 cells (in yellow) emitted virtual rays reaching the
main accumulation region observable in panel (vii) (accumulated ray intensity >3.5). a.u., arbitrary unit. (ii–ix) 2D cross-sectional maps for increasing
depth along the Z-axis. Scale bar =10 µm.
COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z ARTICLE
COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio 5
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Fig. 4 3D organization of typical cellular and subcellular structures. Only 21 entirely contained cells are considered here. aVolumetric correlations
between cytoplasm, nucleus and mitochondrial network. bCell sizes related to distance from blood capillary: (i) Reconstructed 3D image of cells and blood
capillary. Cell colour is related to its volumetric size. (ii–ix) 2D cross-sectional maps for increasing depth along the Z-axis. (i–ix) Blood capillary portions are
in red. Scale bar =10 µm. cCorrelation between distance to blood capillary and cellular/subcellular structures or nucleocytoplasmic ratio. a,cSpearman
correlations. rand pvalues are as indicated in corresponding graph.
ARTICLE COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z
6COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio
Content courtesy of Springer Nature, terms of use apply. Rights reserved

capillary portion (Supplementary Fig. 2, Supplementary Video 1),
a direct correlation between the size of tumour cells, of mito-
chondria and of nuclei (Fig. 4a) and tumour cells polarizing in the
direction of a hot spot corresponding to a bile canaliculus-like
structure (Fig. 3, Supplementary Video 5). Together, these data
suggest that our samples preparation fully preserved the HB
sample ultrastructure following fixation and staining procedures,
and consequently, that our structural observations are reliable.
Our data also demonstrate the feasibility of our methodology to
study the 3D organization of tumoral tissue using specific
bioarchitectural parameters (Fig. 5) and the potential advance it
represents in understanding cancer biology. The information
obtained on the organization of liver cancer cells throws light on
tumour physiology. In particular, it suggests that blood capillaries
could determine the ultrastructure of tumoral tissue by control-
ling the alignment and size of tumour cells and their subcellular
components. However, further studies are required to strengthen
this hypothesis. In addition, we found that a bile canaliculus-like
structure guides the spatial arrangement and the polarity of
tumour cells in HB tissue, as it does in normal liver tissue to
subsume the physiological functions of normal hepatocytes15.
This study is a preliminary attempt to unravel the internal
architecture of tumoral tissue and has limitations. First, it is
difficult to automatically segment-specific structures such as
tumour-infiltrating macrophages and endoplasmic reticulum.
The procedure requires visual control and manual correction,
which are time-consuming. Second, our sample set was small.
Future studies with a larger number and variety of tumoral tissues
would allow us to generate more digitalized samples that in turn
would improve the automatic segmentation process. A larger
dataset would also allow us to better evaluate whether blood
capillaries and bile canaliculus-like structures, which are common
features in HB and more generally in epithelial liver tumours,
influence the polarity and spatial organization of tumour cells and
their organelles. These techniques coupled with current omic
approaches would provide much insight into the relationship
between the structural organization of tumours, and their cellular
functions and metabolic activities.
Future possible directions of onconanotomy-based exploratory
studies are the following: (a) analysing virtually all types of solid
tumours to generate a virtual 3D biorepository whose data would
be particularly valuable in the case of rare tumours or precious
samples such as biopsies; (b) comparing matched tumour samples
such as changes occurring before and after treatment or primary
versus metastatic features, and (c) identifying structural subtypes
and verifying their overlap with different histological groups such
as embryonal versus foetal versus small undifferentiated cell
subtypes described in HB8,16. Since this methodology is adapted
to lab tumour models like cell xenografts, tumour-like spheroids
and organoids, our combined approach could help investigators
to clarify the role of genes and molecular pathways involved in
the architecture of tumoral tissue and to investigate the response
of tumour cells to drugs (e.g. checkpoint inhibitors, monoclonal
antibodies) and treatments (e.g. radiotherapy). Finally, our
approach is perfectly transferable to other biological kingdoms
and fields based on tissue development including for instance
plants, fungi, neuroscience, embryogenesis and biomaterials.
Based on our approach (Fig. 5), 3D EM users could measure
valuable structural parameters (e.g. size, component ratio,
polarity, etc.) and potential additional ones (e.g. cellular density,
surface, texture, etc.) related to tumour tissue components using
biopsies, tumour resections or lab models4. These analyses should
help investigators (1) to better understand how a tumour tissue is
spatially constructed, (2) to determine in a dynamic or com-
parative manner (e.g. at different tumour development periods)
the close relationships between tumour cells and other tissue
Fig. 5 Graphical summary of the 3D EM approach used to study tumour tissue architecture. The figure is divided into three steps: the technical workflow,
the segmented tissue elements and the different bioarchitectural parameters used to analyse the 3D organization of tumour tissue.
COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z ARTICLE
COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio 7
Content courtesy of Springer Nature, terms of use apply. Rights reserved

elements (e.g. blood capillary, nerve endings, immune cells, etc.)
and (3) to study in 3D the elements constituting a “tumour
niche”21. Such investigations could also help to clarify the
mechanisms by which the tissue is expanding while growing and
in which direction, how tumour cells invade the surrounding
tissue by outcompeting with non-tumoral cells and how the
metastatic cells traverse the vascular wall. By performing com-
parative analysis, investigators could evaluate at a nanoscale level
the efficacy of an antitumoral drug transported by the blood
circulation on tumour cells. The implementation of our approach
to histopathology practice may also provide additional bio-
markers for the diagnosis, the prognosis, the drug-response pre-
diction and help to stratify tumours. To sum up, our procedure
should boost the interest of the scientific and clinical commu-
nities working in cancer for 3D EM technologies4.
In conclusion, our study aims to foster the development of
what we have termed onconanotomy, a novel research field in
oncology, that will complement other innovative imaging tech-
nologies such as two-photon excitation microscopy (also known
as 2PEF) and correlative light-electron microscopy (also known
as CLEM). This field could be backed up by the development of
open-source databanks gathering all 3D imaging data at a
worldwide level in order to boost research and promote coop-
eration in cancer, as was previously the case with Gene databanks
and Cell and Tissue atlases. By accessing 3D image databanks,
investigators could upload valuable information and perform
comparative investigations to unravel the ultrastructure of solid
tumours. The potential focuses include cell-cell and cell-
extracellular matrix interactions, intratumoral heterogeneity,
pro-metastatic mechanisms, angiogenesis, drug response, ther-
apeutic efficacy and regulatory factor gradation. We believe that,
in few decades from now and thanks to digital approaches and
artificial intelligence, these approaches could enrich the field of
oncology with the support of other technologies such as single-
cell RNA-seq, microfluidics and complex 3D lab models. It might
especially provide ground-breaking insights into how cells dis-
seminate and how drugs reach the targets cells. A better com-
prehension of the organization of tumour cells and clinical
outcome could help in identifying the key structural parameters
of cancers. In turn, this could lead to substantial advances in both
the understanding of cancers and the treatments of adult and
paediatric cancer patients.
Methods
Ethical clearance. PDX development program was approved by the Hospital ethics
committee, each sample was implanted upon signature of the related informed
consent form. Mice were maintained in specific pathogen-free animal housing at
the Center for Exploration and Experimental Functional Research (CERFE, Evry,
France) animal facility. All experimental procedures related to PDX development
were conducted in accordance with French regulatory legislation concerning the
protection of animals used for scientific purposes were approved by the CERFE
animal facility ethics committee and by the Ministère de l’Agriculture et de l’Ali-
mentation, France.
PDX establishment. At surgery, tumour fragments were sampled from the
resected tumour and placed in cell culture medium supplemented with penicillin/
streptomycin and with or without 5% BSA on ice. Tumour samples were chopped
into 4 × 3 mm fragments and grafted in the interscapular region of 6-8 week-old
female athymic nude mice (Athymic Nude-Foxn1nu, ENVIGO-Harlan Labora-
tories, Gannat, France). Tumour growth from the first implantation occurred with
a delay spanning between 1 and 5 months. Growing tumours were serially trans-
planted onto recipient mice and underwent comparative examination to confirm
preservation of their histological features. To immortalize each PDX, vials of 4 ×
3 mm fragments from tumours at different passages were placed in a solution of
90% FCS/10% DMSO or glycerol, and stored at −150 °C.
PDX samples. PDX was generated in compliance with the informed consent form
signed by the patients and developed as previously described9. The main clinical
and genetic characteristics of HB PDX samples are shown in Supplementary
Table 1.
PDX fragments fixation. For each model, 5 fragments (2 × 1 mm) were taken and
placed in ice-cold fixing solution (2% PFA and 2.5% glutaraldehyde in 0.15 M
Cacodylate +2 mM CaCl
2
buffer). After 3-h fixation with gentle stirring, the
fragments were transferred to a tube containing Cacodylate buffer and kept
overnight at 4 °C before shipment to the Imaging platform.
Serial block-face scanning electron microscopy sample preparation. Tissue was
prepared for SBF-SEM as previously described10. The samples were fixed with a 2%
paraformaldehyde and 2,5% solution of glutaraldehyde in 0.15 M cacodylate buffer
(pH 7.4) for 2 h at room temperature and then were washed 5 × 3 min in cold
0.15 M cacodylate buffer. En bloc contrast staining was performed by consecutive
incubations in heavy metal containing solutions. The first staining step was a 1 h
incubation on ice in 2% OsO
4
containing 1.5% potassium ferrocyanide in 0.15 M
cacodylate buffer. After washing 5 × 3 min in ultrapure water, the samples were
incubated for 20 min in a fresh thiocarbohydrazide solution (1% w/v in water) at
room temperature. The next wash step was followed by incubation in 2% osmium
in water at RT for 30 min and after washing 5 × 3min in ultrapure water 2% uranyl
acetate at 4 °C overnight. The following day, Walton’s lead aspartate staining was
performed for 30 min at 60 °C. For this, a 30 mM L-aspartic acid solution was used
to freshly dissolve lead nitrate (20 mM, pH 5.5), the solution was filtered and blocks
incubated for 30 min at 60 °C. After final washing steps, the samples were dehy-
drated using ice-cold solutions of 30%, 50%, 70%, 90%, 2×100% ethanol (anhy-
drous), 2 × 100% aceton, 10 min each. Resin embedding was done using Epon by
first placing the samples in 25% aceton/Epon for 2 h, 50% aceton/Epon for 2 h, 75%
aceton/epon for 2 h and followed by 2 incubations in 100% epon (overnight, 8 h).
The samples were put in fresh epon’s resin and placed at 60 °C for 48h. Once the
resin block were hardened, they were roughly cut with a razor blade to generate a
pyramid shape and mounted on aluminium specimen pins using a silver filled
conductive resin (Epotek-Delta microscopies, Mauressac, France). After 24 h of
polymerisation at 60 °C, the samples were trimmed with a diamond knife (Dia-
tome, Nidau, Switzerland). Silver filled conductive resin was used to electrically
connect the edges of the tissue to the aluminium pin. The entire sample was then
sputter coated with a 5–10 nm layer of gold to enhance conductivity.
Transmission electron microscopy. The samples were first analysed by TEM to
control morphology and define the region of interest. Ultra-thin sections were cut
and deposited on copper grids and observed with a Hitachi H7650 transmission
electron microscope (Japan).
Serial block-face scanning electron microscopy imaging. The sample on the pin
was placed into the carrier that fits into the3View ultramicrotome (3viewXP2-
Gatan Inc., Pleasanton, CA, USA) on a ZEISS Gemini field emission gun SEM300
(Zeiss - Marly-le-Roi - France). The block face was imaged with an accelerating
voltage at 1.2 kV using the Back scattered electron with a specific BSE Detector (On
point - Gatan Inc., Pleasanton, CA, USA).
Pre-processing and structure segmentation. Image analysis was performed
using three complementary software packages. Digital Micrograph (Gatan Inc.,
Pleasanton, CA, USA) was used to align the images with the Image Alignment
plugin using the combined default filter (which combines soft rectangle and
bandpass filters). The aligned images were saved in Gatan format “dm4”as a single
stack. Fiji software was used to convert the 32-Bit images into 8-Bit images to be
less resource-consuming and facilitate segmentation. Then the image was cropped
to the optimal (larger) rectangular field of view, and the brightness and contrast
were automatically adjusted (Z-scoring on a slice-by-slice basis). Finally, VAST-
Lite software was used for the manual segmentation of the elements of interest
(cells, nuclei and mitochondria)14. Image processing was done following the
pipeline shown in Supplementary Fig. 10. All manual segmentation was done using
an interactive drawing tablet and pen (Cintiq pro 5, Wacom).
The manual segmentation of both cells and nuclei was done only on 10% of the
stack of images (1 image out of 10) using the VAST-Lite software14. These manual
segmentations were propagated to neighbouring slices using a so-called « Optical
Flow » algorithm (see “Source codes”section below) applied on the EM images,
similar to the approach described by Huang T.C. et al.22. This procedure allowed us
to segment partially and entirely 182 tumour cells, 113 nuclei in a relatively decent
time frame (several weeks were mandatory). However, the segmentation of
mitochondria turned out to be a different challenge. The reason is two-fold: (1) the
volume of mitochondria is far lower the size of cells nuclei, preventing the use of
the above-mentioned propagation strategy; (2) the number of mitochondria is
much greater than the number of cells and nuclei, further hampering in turn the
amount of work. Mitochondria were thus segmented manually on one single cell to
feed a deep-learning algorithm (using a 3D U-net architecture23 and an
implementation based on Tensorflow 1.4 and Keras 2.2.4), which could be
subsequently used for semi-automatic segmentation. The 3D U-Net architecture
used for the segmentation of mitochondria is presented in Supplementary Fig. 11.
The architecture is similar to the one proposed by Huo et al.24. We used one single
input channel (i.e., the SBF-image). The loss function was a combination of
categorical cross-entropy and Dice25. The optimizer algorithm was Adam with
default parameters26. 100 epochs were performed and the batch size was equal to 1.
ARTICLE COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z
8COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio
Content courtesy of Springer Nature, terms of use apply. Rights reserved

Due to limited GPU memory, we used a patch-wise strategy27: practically, local
patches of 128 × 128 × 128 voxels were extracted in the training SBF image and
used as inputs for the CNN. To finish, erroneous automatic segmentations of
mitochondria were corrected manually using the VAST-Lite14 and Ilastik13
software.
Analysis of orientation of structures. Each cell, nucleus and capillary portion was
characterized using its main axis. To this end, a Principal Component Analysis
(PCA) was applied on the voxel coordinates within the segmented mask of a
structure of interest. Only tumour cells with a complete nucleus were considered.
The following two analyses were conducted.
Analysis of alignment in 2D plane of cells and nuclei. The main axes of cells were
used to calculate the best 2D “alignment plane”(in the least-squares sense: we
minimized the sum of squared differences between the observed main axes of cells,
and the fitted value provided by a 2D plane equation). Angles between the main
axis of each cell and the alignment plane were calculated. The same analysis was
conducted subsequently for nuclei.
Determination of cell clusters with polarized shape orientation. 3D virtual rays were
emitted by each cell along its main axis in both directions (the ray radius was a
user-defined input parameter for the algorithm and was set to 15 µm, ray intensity
proportional to the Euclidean distance to the main axis, with a value of 1 on the
main axis). To this end, Bresenham’s line algorithm28 was employed and adapted
to our needs. The accumulated beam intensity was calculated on a voxel-by-voxel
basis. Cells emitting virtual rays reaching a specific accumulation region were
subsequently selected. Thus, cells converging toward a common region could be
listed.
Source codes. The code designed to identify the analyzed bioarchitectural para-
meters (alignment, volume, shape orientation) can be found in https://github.com/
bsenneville/Onconanotomy/. The code implementing the above-mentioned optical
flow algorithm employed for the semi-automatic segmentation of cells and nuclei
can be found in https://github.com/bsenneville/2D_Optical_Flow/29. Both codes
were developed under the commercial software Matlab ©1994-2021 The
MathWorks, Inc.
Infographics. Blender version 2.90.1 (https://www.blender.org) was used to create
the illustrations (images and Supplementary videos). The Cycles render engine was
used for all the final renderings. To render polygonal meshes, the data were
imported from VTK files using the BVTK addon (https://github.com/tkeskita/
BVtkNodes). Depending on the data complexity, the meshes were decimated to
lower the polygon count, but still preserving the overall shape and most of the
details. For volumetric rendering, the original stack data was converted from a
multi-page Tiff image file to a downscaled 2D tile map, which was then loaded and
translated to a 3D volume into Blender using a custom shader setup. ImageJ
(https://imagej.nih.gov/ij/), XNview (https://www.xnview.com) and the “montage”
command line tool of ImageMagick (https://imagemagick.org/script/montage.php)
were used for image conversion, downscaling and stitching. Downscaling is
mandatory, since the original amount of data is not always practical to work with
and is not necessary for the purpose of visualization.
Statistical analyses. Statistical analyses were performed using GraphPad Prism
7.05 software. Spearman’s nonparametric correlation was used to compare the size
of different cell components or the size of a specific component to the distance to a
blood capillary. Results were considered significant when p< 0.05.
Reporting summary. Further information on research design is available in the Nature
Research Reporting Summary linked to this article.
Data availability
The data that support the findings of this study are available within the paper and
its Supplementary information files.
Received: 17 August 2021; Accepted: 12 November 2021;
References
1. Indersie, E. et al. MicroRNA therapy inhibits hepatoblastoma growth in vivo
by targeting beta-catenin and Wnt signaling. Hepatol. Commun. 1, 168–183
(2017).
2. Cartier, F. et al. New tumor suppressor microRNAs target glypican-3 in
human liver cancer. Oncotarget 8, 41211–41226 (2017).
3. Denk, W. & Horstmann, H. Serial block-face scanning electron microscopy to
reconstruct three-dimensional tissue nanostructure. PLoS Biol. 2, e329 (2004).
4. Riesterer, J. L. et al. A workflow for visualizing human cancer biopsies using
large-format electron microscopy. Methods Cell Biol. 158, 163–181 (2020).
5. Knott, G., Marchman, H., Wall, D. & Lich, B. Serial section scanning electron
microscopy of adult brain tissue using focused ion beam milling. J. Neurosci.
28, 2959–2964 (2008).
6. Titze, B. & Genoud, C. Volume scanning electron microscopy for imaging
biological ultrastructure. Biol. Cell 108, 307–323 (2016).
7. Indersie, E. et al. Tracking cellular and molecular changes in a species-specific
manner during experimental tumor progression in vivo. Oncotarget 9,
16149–16162 (2018).
8. Hooks, K. B. et al. New insights into diagnosis and therapeutic options for
proliferative hepatoblastoma. Hepatology 68,89–102 (2018).
9. Nicolle, D. et al. Patient-derived mouse xenografts from pediatric liver cancer
predict tumor recurrence and advise clinical management. Hepatology 64,
1121–1135 (2016).
10. Deerinck, T. J., Bushong, E., Thor, A. & Ellisman, M. H. NCMIR methods for
3D EM: a new protocol for preparation of biological specimens for serial
blockface scanning electron microscopy. Microscopy 1,6–8 (2010).
11. Dekmezian, R., Sneige, N., Popok, S. & Ordonez, N. G. Fine-needle aspiration
cytology of pediatric patients with primary hepatic tumors: a comparative
study of two hepatoblastomas and a liver-cell carcinoma. Diagn. Cytopathol. 4,
162–168 (1988).
12. Pietsch, T. et al. Characterization of the continuous cell line HepT1 derived
from a human hepatoblastoma. Lab Invest 74, 809–818 (1996).
13. Berg, S. et al. ilastik: interactive machine learning for (bio)image analysis. Nat.
Methods 16, 1226–1232 (2019).
14. Berger, D. R., Seung, H. S. & Lichtman, J. W. VAST (volume annotation and
segmentation tool): efficient manual and semi-automatic labeling of large 3d
image stacks. Front Neural Circuits 12, 88 (2018).
15. Cullen, J. M. & Stalker, M. J. in Jubb, Kennedy & Palmer’s Pathology of
Domestic Animals Vol. 2 (ed Grant Maxie, M.) 258–352.e251 (2016).
16. Lopez-Terrada, D. et al. Towards an international pediatric liver tumor
consensus classification: proceedings of the Los Angeles COG liver tumors
symposium. Mod. Pathol. 27, 472–491 (2014).
17. Chan, Y. H. & Marshall, W. F. Scaling properties of cell and organelle size.
Organogenesis 6,88–96 (2010).
18. Schmoller, K. M. The phenomenology of cell size control. Curr. Opin. Cell
Biol. 49,53–58 (2017).
19. Fuhrmann, D. C., Wittig, I., Heide, H., Dehne, N. & Brune, B. Chronic
hypoxia alters mitochondrial composition in human macrophages. Biochim.
Biophys. Acta 1834, 2750–2760 (2013).
20. Lages, Y. M. et al. Low oxygen alters mitochondrial function and response to
oxidative stress in human neural progenitor cells. PeerJ 3, e1486 (2015).
21. Laplane, L., Duluc, D., Larmonier, N., Pradeu, T. & Bikfalvi, A. The multiple
layers of the tumor environment. Trends Cancer 4, 802–809 (2018).
22. Huang, T. C. et al. Semi-automated CT segmentation using optic flow and
Fourier interpolation techniques. Comput Methods Prog. Biomed. 84, 124–134
(2006).
23. Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional Networks for
Biomedical Image Segmentation. In International Conference on Medical
Image Computing and Computer-Assisted Intervention 234–241 (Springer,
Cham, 2015) https://doi.org/10.1007/978-3-319-24574-4_28.
24. Huo, Y. et al. 3D whole brain segmentation using spatially localized atlas
network tiles. Neuroimage 194, 105–119 (2019).
25. Sugino, T. et al. Loss weightings for improving imbalanced brain structure
segmentation using fully convolutional networks. Healthcare (Basel) 9,https://
doi.org/10.3390/healthcare9080938 (2021).
26. Kingma, D. P. & Ba, J. Adam: A Method for Stochastic Optimization. CoRR
abs/1412.6980 (2015).
27. Liu, Y., Ren, Q. R., Geng, J. H., Ding, M. & Li, J. Y. Efficient patch-wise
semantic segmentation for large-scale remote sensing images. Sensors-Basel
18, 3232 (2018).
28. Bresenham, J. Algorithm for computer control of a digital plotter. IBM Syst. J.
4,25–30 (1965).
29. Zachiu, C., Papadakis, N., Ries, M., Moonen, C. & Denis de Senneville, B. An
improved optical flow tracking technique for real-time MR-guided beam
therapies in moving organs. Phys. Med Biol. 60, 9003–9029 (2015).
Acknowledgements
This work was supported by the charity Eva pour la Vie, La Fondation ARC pour la
Recherche sur le Cancer (contract N° PJA 20191209631), La Région Nouvelle-Aquitaine,
La Fondation Groupama pour la Santé and Groupama Centre-Atlantique. Microscopy
Imaging was performed at the Bordeaux Imaging Centre, which is a member of the
FranceBioImaging national infrastructure (ANR-10-INBS-04). Computational analysis
was carried out using the PlaFRIM experimental testbed, supported by Inria, CNRS
(LABRI and IMB), Université de Bordeaux, Bordeaux INP and Conseil Régional
COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z ARTICLE
COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio 9
Content courtesy of Springer Nature, terms of use apply. Rights reserved

d’Aquitaine (see https://www.plafrim.fr/). Computer time for this study was provided by
the computing facilities at MCIA (Mésocentre de Calcul Intensif Aquitain) of the Uni-
versité de Bordeaux and of the Université de Pau et des Pays de l’Adour.
Author contributions
Fresh paediatric tumour fragments were provided by S.B. and C.C. HB-PBX were
developed by XenTech company and provided by K.F. and S.C. Image acquisition and
analyses were performed by M.B., F.Z.K. and E.G. The manual segmentations and clean-
up of tissue and cellular components were performed by F.Z.K., M.B. and C.F.G. B.D.d.S.
performed the mathematical and computational analyses. A.L. generated 3D images and
videos. C.F.G., S.C. and E.G. supervised the work. C.F.G. obtained the financial grants.
All authors actively participated in the writing of the manuscript.
Competing interests
The authors declare no competing interests.
Additional information
Supplementary information The online version contains supplementary material
available at https://doi.org/10.1038/s42003-021-02919-z.
Correspondence and requests for materials should be addressed to Etienne Gontier or
Christophe F. Grosset.
Peer review information Communications Biology thanks the anonymous reviewers for
their contribution to the peer review of this work. Primary Handling Editors: Chao Zhou
and Karli Montague-Cardoso.
Reprints and permission information is available at http://www.nature.com/reprints
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative
Commons license, and indicate if changes were made. The images or other third party
material in this article are included in the article’s Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not included in the
article’s Creative Commons license and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this license, visit http://creativecommons.org/
licenses/by/4.0/.
© The Author(s) 2021
ARTICLE COMMUNICATIONS BIOLOGY | https://doi.org/10.1038/s42003-021-02919-z
10 COMMUNICATIONS BIOLOGY | (2021) 4:1390 | https://doi.org /10.1038/s42003-021-02919-z | www.nature.com/commsbio
Content courtesy of Springer Nature, terms of use apply. Rights reserved

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com