Content uploaded by Sean M. Hacking
Author content
All content in this area was uploaded by Sean M. Hacking on Aug 12, 2021
Content may be subject to copyright.
Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=iusp20
Ultrastructural Pathology
ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/iusp20
Deep learning for the classification of medical
kidney disease: a pilot study for electron
microscopy
Sean Hacking & Vanesa Bijol
To cite this article: Sean Hacking & Vanesa Bijol (2021): Deep learning for the classification
of medical kidney disease: a pilot study for electron microscopy, Ultrastructural Pathology, DOI:
10.1080/01913123.2021.1882628
To link to this article: https://doi.org/10.1080/01913123.2021.1882628
Published online: 14 Feb 2021.
Submit your article to this journal
Article views: 82
View related articles
View Crossmark data
Deep learning for the classication of medical kidney disease: a pilot study for
electron microscopy
Sean Hacking and Vanesa Bijol
Department of Pathology and Laboratory Medicine, Donald and Barbara Zucker School of Medicine at Northwell, Manhasset, New York, USA
ABSTRACT
Articial intelligence (AI) is a new frontier and often enigmatic for medical professionals. Cloud
computing could open up the eld of computer vision to a wider medical audience and deep
learning on the cloud allows one to design, develop, train and deploy applications with ease. In the
eld of histopathology, the implementation of various applications in AI has been successful for
whole slide images rich in biological diversity. However, the analysis of other tissue medias,
including electron microscopy, is yet to be explored. The present study aims to evaluate deep
learning for the classication of medical kidney disease on electron microscopy images: amyloido-
sis, diabetic glomerulosclerosis, membranous nephropathy, membranoproliferative glomerulone-
phritis (MPGN), and thin basement membrane disease (TBMD). We found good overall classication
with the MedKidneyEM-v1 Classier and when looking at normal and diseased kidneys, the average
area under the curve for precision and recall was 0.841. The average area under the curve for
precision and recall on the disease only cohort was 0.909. Digital pathology will shape a new era for
medical kidney disease and the present study demonstrates the feasibility of deep learning for
electron microscopy. Future approaches could be used by renal pathologists to improve diagnostic
concordance, determine therapeutic strategies, and optimize patient outcomes in a true clinical
environment.
ARTICLE HISTORY
Received 30 December 2020
Accepted 25 January 2021
KEYWORDS
Deep Learning; Medical
Kidney Disease; Electron
Microscopy; Cloud
Computing; 21st Century
Medicine
Introduction
Today, the field of artificial intelligence (AI) is
exploding, primarily due to deep learning (DL),
a subset of machine learning (ML) which utilizes
artificial neural networks, modeled after the occipi-
tal cortex of the human brain.
1
This excitement has led to the development of
numerous computational models with utility for his-
tological segmentation of and classification of glo-
meruli from mouse and human kidney biopsies.
2–18
One study was found to have the ability to predict
clinical phenotype utilizing convolutional neural
networks (CNN) and the assessment of kidney
fibrosis.
18
Current state of the art is based on hema-
toxylin and eosin (H&E), periodic acid Schiff (PAS)
and trichrome-stained tissues; however, there are
limited data on electron microscopy (EM) tissue
media and computer vision.
The first convolutional neural network for multi-
class segmentation was validated on PAS-stained
nephrectomy samples and transplant biopsies.
5
Despite recent advancements, models like these
really have no utility for clinical practice. Often
termed “low level,” these approaches do not per-
form diagnostics and guide treatments for patients
with medical kidney disease.
4
More recently, Ginley et al. developed a digital
pipeline to classify renal biopsies from patients with
diabetic nephrosclerosis (DN).
16
However, this
model did not include many other medical nephro-
pathies, such as amyloidosis, fibrillary glomerulo-
nephritis, membranous nephropathy (MN), thin
basement membrane disease (TBMD), postinfec-
tious glomerulonephritis (PIGN), and membrano-
proliferative glomerulonephritis (MPGN).
Creating a computation model capable of robust
disease classification is the next frontier for compu-
tational image analysis in digital nephropathology.
With recent developments in cloud computing,
models can be developed anywhere and application
programming interfaces (API) can be uploaded
from the cloud onto mobile or desktop-based plat-
forms. The advantages for using cloud computing
CONTACT Sean Hacking shacking@northwell.edu Department of Pathology and Laboratory Medicine, Donald and Barbara Zucker School of Medicine
at Northwell, Manhasset, New York, USA.
ULTRASTRUCTURAL PATHOLOGY
https://doi.org/10.1080/01913123.2021.1882628
© 2021 Taylor & Francis Group, LLC
in the deep learning environment are that it allows
large datasets to be easily integrated and organized
for model training; it also allows models to scale
efficiently and at a lower cost. The cloud leverages
the advantages associated with distributing net-
works and by deep learning on the cloud, one can
design, develop, and train deep learning applica-
tions easily.
The evaluation of ultrastructural features is often
performed in medical practice and electron micro-
scopy (EM) is the champion of this field.
19
Recent
methods for partial scanning transmission electron
microscopy (TEM) and deep learning were able to
achieve high performance with adaptive learning,
rate clipping of loss spikes, in addition to an aux-
iliary trainer network,
3
while other groups have
been able to utilize machine learning on small sub-
sets of labeled EM data.
2
In the field of renal pathol-
ogy, EM images are often very helpful for more
precise diagnosis and better characterization of
the disease processes. This begs the question: can
deep learning be utilized on EM images for the
diagnosis of medical kidney disease?
As tissue analysis is performed on more than
traditional hematoxylin and eosin (H&E) stained
tissue, this study also serves to expand the use
of deep learning to other tissue media. The use
of immunofluorescence (IF) and immunohisto-
chemistry (IHC) is ubiquitous to the practice of
pathology and methods such as in-vivo micro-
scopy are also becoming more understood.
20–22
This suggests that it may be valuable to validate
deep learning on various visual mediums for
future use in clinical practice.
This publication is the first to evaluate deep
learning and EM for the purpose of classifying
medical kidney disease. Herein, we developed the
MedKidneyEM-v1 Classifier, which was derived
from conventional digital EM images. The present
study evaluates deep learning in EM, while also
stimulating discussion on cloud computing and
the practice of 21
st
century medicine.
Materials and methods
Ethics statement
Experimental protocols were approved by
Institutional Review Board at the Office of the
Human Research Protection Program licensing
committee. All methods were carried out in accor-
dance with relevant guidelines and regulations.
Patient consent was not required by the institutional
review board (IRB) committee due to the retrospec-
tive nature of the study (Northwell Health IRB num-
ber: 20–0824).
Study design
This was a retrospective study of medical kidney dis-
ease patients diagnosed at the Northshore University
Hospital and Long Island Jewish Medical Center,
Northwell Health. Diagnoses were made based on
the clinical findings in addition to the assessment of
hematoxylin and eosin (H&E), periodic acid-Schiff
(PAS), Jones methenamine, and trichrome-stained
slides, EM findings, and immunofluorescence (IF).
For the purpose of computation in this current
study, EM images were classified for training, valida-
tion and testing based on the final diagnosis made by
an expert renal nephropathologist (VB) in collabora-
tion with clinical nephrologists and other consultant
specialists as necessary.
Transmission electron microscopy
EM was to allow for direct detection followed by
negative staining as originally proposed by Brenner
and Horne (1959) (Figure 1).
23
Tissues were fixed with
2–2.5% glutaraldehyde in 0.1 phosphate buffer (pH
7.0 − 7.3) and examined on a JEOL JEM-100CX II
with a tungsten-filament 100kV transmission electron
microscope. Digital images were performed with an
Advanced Microscopy Techniques (AMT) camera
(model: 1412AM-T1-FW-AM), Woburn, MA.
Ultrathin sections were examined using standard
techniques. Morphological features of all glomerular
structures were evaluated, including parietal and visc-
eral epithelial cells, glomerular basement membrane
thickness and texture, endothelial cell morphology, as
well as mesangial space with cells and matrix. The
presence of electron-dense deposits and their distribu-
tion and substructural organization was also noted.
Classication of medical kidney disease
EM images were collected from patients diagnosed
with the following disease entities: amyloidosis,
2S. HACKING AND V. BIJOL
membranous nephropathy (MN), membranoproli-
ferative glomerulonephritis (MPGN), thin basement
membrane disease (TBMD), diabetic glomerulo-
sclerosis, and normal control kidneys.
Amyloid deposits on EM reveal fibrillary sub-
structural organization, with randomly arranged
non-branching fibrils 8–12 nm in diameter. MN is
characterized by the presence of subepithelial elec-
tron-dense deposits and intervening basement mem-
brane spikes, with associated diffuse effacement of
podocyte foot processes. In early stages of MN, sub-
epithelial dense deposits are sparsely distributed and
basement membrane spikes are not well established.
In membranoproliferative pattern of glomerular
injury, glomerular capillary walls become remo-
deled due to the presence of subendothelial dense
deposits and cellular interposition; a new basement
membrane material is often laid down under the
displaced endothelium resulting in double contour
formation.
Glomerular basement membranes can be attenu-
ated due to an inherited abnormality of basement
membrane collagens. Thin glomerular basement
membranes are of normal structure but when mea-
sured using standard intercept methods, they are
less than 264 nm in diameter. Splintering of the
basement membranes is seen in more severe defects
that are considered under the hereditary nephritis
and those changes should not be seen in a thin
glomerular basement membrane lesion.
Diabetic nodular glomerulosclerosis occurs
in patients with the clinical picture of diabetic
nephropathy. On EM, the mesangium is expanded
by matrix and cells, often with deposits of cellular
debris dispersed in the matrix. Glomerular
basement membranes are irregularly thickened.
Examples of different EM findings are available in
Figure 2.
Deep learning on the cloud
The MedKidneyEM-v1 Classifier was developed on
the Google Cloud Platform and was built on the
AutoML Vision application programming inter-
face. Cloud AutoML had the advantage of being
fully integrated with other Google Cloud services,
which was ideal for storing and training data in
Cloud Storage and sharing of APIs.
24
Google
Cloud AutoML Vision relies on the principles of
deep learning; however, it also utilizes neural archi-
tecture searching. This transfer learning is used
under the hood to find the best deep learning net-
work architecture with optimal hyperparameter
configuration that minimizes the loss function for
the model.
24
A total of 21 node hours were used for
the creation of this computational assay. This
Figure 1. Negative-stain transmission electron microscop
ULTRASTRUCTURAL PATHOLOGY 3
MedKidney-v1 Classifier is also fully connected,
meaning that it goes directly from image to disease
classification and the ability to be deployed to both
desktop and mobile platforms. An overview of the
pipeline for our classifier is available in Figure 3.
Statistical analysis
Analyses were planned before the evaluations in the
training, validation and testing cohorts. The prede-
fined primary analysis for the MedKidneyEM-v1
Classifier was the creation of precision-recall trade-
off curves for different medical kidney diseases.
Confusion matrices were also classified to show
how often the model classified each disease cor-
rectly, and which disease were most often confused
for other entities. All analytical calculations were
calculated on Google Cloud Platform for
a confidence threshold of 0.5.
Results
This study included an image database with 600
images from normal kidney (80) and the following
medical kidney diseases: amyloidosis (93), diabetic
glomerulosclerosis (106), MPGN (91), MN (120)
and TBMD (110). The classifier was trained on
a cohort consisting of 74 amyloid, 84 diabetic glomer-
ulosclerosis, 72 MPGN, 96 MN, 64 normal, and 87
TBMD medical kidney EM images. An independent
validation cohort consisted of 8 amyloidosis, 10 dia-
betic glomerulosclerosis, 9 MPGN, 12 MN, 8 normal,
and 10 TBMD. The classifier was tested on 11 amy-
loidosis, 12 diabetic glomerulosclerosis, 10 MPGN, 12
MN, 8 normal, and 13 TBMD. Magnification for
images was as follows: amyloidosis (2700–100000x),
diabetic glomerulosclerosis (4000–27000x), MPGN
(4000–27000x), MN (6700–27000x), normal (5000–
14000x), and TBMD (6700–20000x). Case numbers
and image magnification specifics can be found in
Table 1.
Overall performance
The average area under the curve (AUC) for preci-
sion and recall on both normal and diseased kid-
neys was 0.841. For a confidence threshold of 0.5,
the precision was 79.63% and the recall 65.15%. The
average area under the curve for precision and
recall on the disease only kidney cohort was 0.909.
Figure 2. Electron microscopy images and the spectrum of medical kidney disease. MPGN, membranoproliferative glomerulonephritis;
TBMD, thin basement membrane disease.
4S. HACKING AND V. BIJOL
For a confidence threshold of 0.5, the precision was
88.89% and the recall 68.97%. Results from AUC
(precision vs recall) and confusion matrix percen-
tages are illustrated in Figure 4.
Disease-specic performance
The normal and disease model had the following
precision values with a confidence threshold of 0.5.
For amyloidosis, precision was 100%, with a recall of
72.73%. For diabetic glomerulosclerosis precision
was 88.89% and the recall 66.67%. For MN, preci-
sion was 77.78% and the recall 58.33%. For MPGN,
precision was 77.78% and the recall 70%. For normal
kidney biopsies, precision was 50% and recall 25%.
For TBMD, precision was 73.33% and recall 84.62%.
The disease only model had the following precision
values with a confidence threshold of 0.5. For amyloi-
dosis, precision was 100% and recall 72.73%. For
diabetic glomerulosclerosis, precision was 100% and
Figure 3. Deep learning on the cloud allows you to design, develop and train applications.
Table 1. Electron microscopy image case numbers and magnification specifics. EM, Electron microscopy; MPGN, Membranoproliferative
glomerulonephritis; TBMD, Thin basement membrane disease.
EM Images Amyloid Diabetes MPGN Membranous Normal TBMD
Total 93 106 91 120 80 110
Training 74 84 72 96 64 87
Validation 8 10 9 12 8 10
Test 11 12 10 12 8 13
Magnification 2700–100000x 4000–27000x 4000–27000x 6700–27000x 5000–14000x 6700–20000x
ULTRASTRUCTURAL PATHOLOGY 5
recall 50%. For MN, precision was 80% with a recall of
66.67%. For MPGN, precision was 75% with a recall
of 60%. For TBMD, precision was 92.31% and recall
92.31%. Results from AUC (precision vs recall) for
individual disease entities are illustrated in Figure 5.
Discussion
The diagnosis of medical kidney disease is based off
the visual assessment of conventional light micro-
scopy, IF, and EM. This is performed by
a consultant renal pathologist, who correlates
pathological findings with the patients’ clinical
profile.
Current approaches may not fully capture the
complex structural and cellular changes present in
medical kidney disease. Saying this, the hope is that
computational pathology will usher in a new era for
nephropathology; moving forward, we will need to
borrow concepts from biology and develop a new
ecosystem. In this study, we build on recent
advancements in deep learning and cloud
Figure 4. Overall performance of the MedKidneyEM-v1Classifier. MPGN, membranoproliferative glomerulonephritis; TBMD, thin
basement membrane disease.
6S. HACKING AND V. BIJOL
computing to develop a classifier (MedKidneyEM-
v1) for medical kidney disease based on electron
microscopy images.
Deep learning has accomplished many recent
achievements for the detection of different diseases
and cancer subtypes,
25
with a number of different
computational models being produced and pub-
lished over the past 20 years. Today, approaches
are beginning to be developed for new visual media
in the field of medicine. An example being the
electrocardiogram (ECG), where deep learning
has been found to provide prognostic information
to the interpretation of 12-lead resting ECGs.
26
There have been numerous recent studies using
AI for medical kidney diseases
2–16
, 8 of them uti-
lized PAS staining, 4 trichrome, while one used
desmin and another H&E.
4
Distinctive from our
study, these studies were performed on whole
slide images (WSI) from mice or human kidney
biopsies. Prior to this publication, the utility of
EM for diagnosis by computational analysis was
largely unknown.
Previous publications showed utility primarily
for segmentation of histological primitives, includ-
ing interstitial fibrosis, tubular atrophy, global glo-
merulosclerosis, glomeruli, Bowman’s capsule,
proximal and distal tubules, as well as the intersti-
tial space and capillaries. Compared to WSIs, EM
images have much more granularity and despite the
lack of color, these images harbor complex visual
profiles. This is often termed: high dimensionality
feature space, something certainly true for EM
images, which actually contain numerous high
dimensional features.
An interesting aspect for discussion is that prior
studies have also been able to provide prognostic
significance as a signature. This was demonstrated
by Kolachalama et al. on 171 WSI from kidney
biopsies based on the assessment of fibrosis.
18
In
theory, EM can be used to optimize diagnostic and
prognostic accuracy in medical kidney disease.
Here, machine learning could increase diagnostic
accuracy, streamline workflows and standardize
diagnostics. The ultimate goal of this being simple:
to improve patient outcomes in a true clinical
setting.
In future, applications in artificial intelligence
could be also used to determine treatments. For
example, in membranous nephropathy, treatment
is based on managing underlying disease states; how-
ever, around 40% of patients recover while others
progress to treatment with immunosuppressive
therapy.
27
With AI, patients at risk for progression
could be identified earlier; while those with good
predictive outcomes could avoid overtreatment.
The use of AI in digital nephropathology should
not be limited to WSI. This is supported by the
findings in our study, where we were able to pro-
duce a computational model with utility for classi-
fication of medical kidney diseases. In doing so, we
were able to achieve good overall classification with
the Google Cloud Platform. When looking at
Figure 5. Area under the curve for precision and recall by individual disease entity. MPGN, membranoproliferative glomerulonephritis;
TBMD, Thin basement membrane disease.
ULTRASTRUCTURAL PATHOLOGY 7
normal and diseased kidneys, the average area
under the curve was 0.841, while the average area
under the curve for precision and recall on the
disease only the kidney cohort was 0.909.
When looking at the classification of normal and
diseased kidneys, the model did misclassify normal
kidneys as TBMD. This was certainly a pitfall which
could lead to problems in clinical practice.
Regarding findings on EM, the mean width of the
GBM in TBMD is 225 ± 20 nm, while in normal
kidneys, it is 354 ± 42 nm.
28
It will be important for
the renal community to understand the limits of
different topologies in computational pathology. At
the same time, we as medical professionals need to
invest time and energy into fostering these technol-
ogies, in order to facilitate uptake for improved
patient care.
Pitfalls also highlight the need for integrated
diagnostics and diagnosis should be based on an
amalgamation of both clinical and pathological
findings, in conjunction with computation. This
has been termed the “digital ecosystem,”
4
where
a visual assessment is performed by a pathologist
and integrated with other domains, including arti-
ficial intelligence, for optimal patient care.
In this publication, EM images were taken at
varying magnifications in order to produce
a robust kidney disease classifier. In cases of amy-
loidosis, for example, image magnification varied
almost 50-fold from 2700 to 100000x. Future clas-
sifiers built on larger image databases, which are
trained at variable magnifications will be important
for bringing this technology to market.
At the time of this publication, 13 image reposi-
tories exist for kidney disease from universities
including the University of Michigan, the
University of Heidelberg, and the University of
Pennsylvania, to name a few.
4
Currently, these
databases are primarily comprised of WSI; how-
ever, different visual media in renal pathology will
grow so the digital pathology repositories would
include EM images.
It is important to mention that IF is integral to
diagnostic renal pathology
29
where characteristic
staining patterns for immunoglobulins and comple-
ment components can be used to diagnose many
forms of glomerulonephritis.
30
One study on pros-
tatectomy specimens developed a deep learning-
based artificial intelligence model to automate the
analysis of IF staining for anti-Ki-67 and ERG
antibodies.
31
They found that AI generated more
accurate and reproducible prediction models for
recurrence and metastasis after surgery, when com-
pared to manual analysis of IF.
31
Considering that IF
is ubiquitous to the field of nephropathology, future
applications should also integrate this technique.
In summary, we developed a clinically useful
medical kidney disease classifier based on deep
learning for EM images. This classifier was validated
in a large patient population, correlating well with
diagnoses given by expert renal pathologist. This
could be particularly useful in resource poor settings
and for institutions lacking subspecialty expertise.
Looking forward, it is unlikely that AI will replace
the consultant role of a renal pathologist, although it
will be a useful tool. Approaches for prognostic
profiling and therapeutic decision-making would
be particularly useful for patients with medical kid-
ney disease.
Availability of Application Programming
Interface
The MedKidneyEM-v1 Classifier can be made available from
the corresponding author upon reasonable request.
Compliance with Ethical Standards
The publication was carried out in accordance with the ethical
standards of the institutional research committee and with the
1964 Helsinki declaration and its later amendments or com-
parable ethical standards.
Disclosure statement
The authors report no conflicts of interest in this work.
Funding
No funding was provided for the production of this
manuscript.
ORCID
Sean Hacking http://orcid.org/0000-0001-6061-0781
8S. HACKING AND V. BIJOL
References
1. Arel I, Rose DC, Karnowski TP: Deep machine learning
- A new frontier in artificial intelligence research
[Research frontier]. IEEE Computational Intelligence
Magazine 2010, 5:13–18.
2. Weber GH, Ophus C, Ramakrishnan L. Automated
labeling of electron microscopy images using deep
learning. 2018 IEEE/ACM Mach Learn HPC Environ
(MLHPC). 2018;1:26–36.
3. Ede JM. Beanland R: partial scanning transmission elec-
tron microscopy with deep learning. Sci Rep.
2020;10:8332. doi:10.1038/s41598-020-65261-0.
4. Barisoni L, Lafata KJ, Hewitt SM, Madabhushi A,
Balis UGJ. Digital pathology and computational
image analysis in nephropathology. Nat Rev
Nephrol. 2020;16:669–685. doi:10.1038/s41581-020-
0321-6.
5. Hermsen M, de Bel T, den Boer M, et al. Deep learning–
Based histopathologic assessment of kidney tissue. J Am
Soc Nephrol. 2019;30:1968–1979. doi:10.1681/
ASN.2019020144.
6. Lutnick B, Ginley B, Govind D, et al. An integrated
iterative annotation technique for easing neural net-
work training in medical image analysis. Nat Mach
Intell. 2019;1:112–119. doi:10.1038/s42256-019-
0018-3.
7. Samsi S, Jarjour WN, Krishnamurthy A: Glomeruli
segmentation in H&E stained tissue using perceptual
organization. 2012 IEEE Signal Processing in Medicine
and Biology Symposium (SPMB), New York, NY, USA.
2012. pp. 1–5.
8. Kato T, Relator R, Ngouv H, et al. Segmental HOG: new
descriptor for glomerulus detection in kidney micro-
scopy image. BMC Bioinformatics. 2015;16:316.
doi:10.1186/s12859-015-0739-1.
9. Gadermayr M, Dombrowski A-K, Klinkhammer BM,
Boor P, Merhof D. CNN cascades for segmenting sparse
objects in gigapixel whole slide images. Comput Med
Imaging Graph. 2019;71:40–48. doi:10.1016/j.
compmedimag.2018.11.002.
10. Gupta L, Klinkhammer BM, Boor P, Merhof D,
Gadermayr M. Iterative learning to make the most of
unlabeled and quickly obtained labeled data in histol-
ogy. Proceedings of the 2nd international conference on
medical imaging with deep learning. In: Cardoso MJ,
Aasa F, Ben G, et al., eds. Proceedings of Machine
Learning Research: PMLR. London, United Kingdom.
2019:215–224.
11. Bukowy JD, Dayton A, Cloutier D, et al. Region-
based convolutional neural nets for localization of
glomeruli in trichrome-stained whole kidney
sections. J Am Soc Nephrol. 2018;29:2081–2088.
doi:10.1681/ASN.2017111210.
12. Gadermayr M, Gupta L, Appel V, Boor P,
Klinkhammer BM, Merhof D. Generative adversarial
networks for facilitating stain-independent supervised
and unsupervised segmentation: a study on kidney
histology. IEEE Trans Med Imaging.
2019;38:2293–2302. doi:10.1109/TMI.2019.2899364.
13. Gadermayr M, Eschweiler D, Jeevanesan A,
Klinkhammer BM, Boor P, Merhof D. Segmenting
renal whole slide images virtually without training
data. Comput Biol Med. 2017;90:88–97. doi:10.1016/j.
compbiomed.2017.09.014.
14. Marsh JN, Matlock MK, Kudose S, et al. Deep learning
global glomerulosclerosis in transplant kidney frozen
sections. IEEE Trans Med Imaging. 2018;37:2718–2728.
doi:10.1109/TMI.2018.2851150.
15. Kannan S, Morgan LA, Liang B, et al. Segmentation of
glomeruli within trichrome images using deep learning.
Kidney Int Rep. 2019;4:955–962. doi:10.1016/j.
ekir.2019.04.008.
16. Ginley B, Lutnick B, Jen KY, et al. Computational
segmentation and classification of diabetic
glomerulosclerosis. J Am Soc Nephrol.
2019;30:1953–1967. doi:10.1681/ASN.2018121259.
17. Stanworth SJ, New HV, Apelseth TO, et al. Effects of the
COVID-19 pandemic on supply and use of blood for
transfusion. Lancet Haematol. 2020;7:e756–e764.
doi:10.1016/S2352-3026(20)30186-1.
18. Kolachalama VB, Singh P, Lin CQ, et al. Association of
pathological fibrosis with renal survival using deep
neural networks. Kidney Int Rep. 2018;3:464–475.
doi:10.1016/j.ekir.2017.11.002.
19. Koster A, Ziese U, Verkleij AJ, Janssen AH, De Jong K.
Three-dimensional transmission electron microscopy:
a novel imaging and characterization technique with
nanometer scale resolution for materials science.
J Phys Chem B. 2000;1:104.
20. Wells WA, Thrall M, Sorokina A, et al. In vivo and ex
vivo microscopy: moving toward the integration of
optical imaging technologies into pathology practice.
Arch Pathol Lab Med. 2019;143:288–298. doi:10.5858/
arpa.2018-0298-RA.
21. Im K, Mareninov S, Diaz MFP, Yong WH. An intro-
duction to performing immunofluorescence staining.
Methods Mol Biol. 2019;1897:299–311.
22. Kim S-W, Roh J, Park C-S. Immunohistochemistry
for pathologists: protocols, pitfalls, and tips.
J Pathol Transl Med. 2016;50:411–418. doi:10.4132/
jptm.2016.08.08.
23. Brenner S, Horne RW. A negative staining method for
high resolution electron microscopy of viruses. Biochim
Biophys Acta. 1959;34:103–110. doi:10.1016/0006-
3002(59)90237-9.
24. Bisong E. Google AutoML: Cloud Vision. New York,
USA/Apress. 2019:581–598.
25. Huang S, Yang J, Fong S, Zhao Q. Artificial intelligence
in cancer diagnosis and prognosis: opportunities and
challenges. Cancer Lett. 2020;471:61–71. doi:10.1016/j.
canlet.2019.12.007.
ULTRASTRUCTURAL PATHOLOGY 9
26. Raghunath S, Ulloa Cerna AE, Jing L, et al.
Prediction of mortality from 12-lead electrocardio-
gram voltage data using a deep neural network. Nat
Med. 2020;26:886–891. doi:10.1038/s41591-020-
0870-z.
27. van den Brand JAJG, van Dijk PR, Hofstra JM,
Wetzels JFM. Long-term outcomes in idiopathic mem-
branous nephropathy using a restrictive treatment
strategy. J Am Soc Nephrol. 2014;25:150–158.
doi:10.1681/ASN.2013020185.
28. Sugiyama N. A clinical and morphological study of thin
basement membrane disease (TBMD). Nihon Jinzo
Gakkai Shi. 1994;36:1010–1020.
29. Singh G, Singh L, Ghosh R, Nath D, Dinda AK.
Immunofluorescence on paraffin embedded renal biop-
sies: experience of a tertiary care center with review of
literature. World J Nephrol. 2016;5:461–470.
doi:10.5527/wjn.v5.i5.461.
30. McCluskey RT. The value of immunofluorescence in
the study of human renal disease. J Exp Med.
1971;134:242–255. doi:10.1084/jem.134.3.242.
31. de la Calle CM, Nguyen HG, Hosseini-Asl E, et al.
Artificial intelligence for streamlined
immunofluorescence-based biomarker discovery in
prostate cancer. J Clin Oncol. 2020;38:279. doi:10.1200/
JCO.2020.38.6_suppl.279.
10 S. HACKING AND V. BIJOL