Machine learning: from radiomics to discovery and routine

Abstract

Machine learning is rapidly gaining importance in radiology. It allows for the exploitation of patterns in imaging data and in patient records for a more accurate and precise quantification, diagnosis, and prognosis. Here, we outline the basics of machine learning relevant for radiology, and review the current state of the art, the limitations, and the challenges faced as these techniques become an important building block of precision medicine. Furthermore, we discuss the roles machine learning can play in clinical routine and research and predict how it might change the field of radiology.

Full text

Review
Radiologe
https://doi.org/10.1007/s00117-018-0407-3
© The Author(s) 2018
G. Langs1·S.Röhrich·J.Hofmanninger·F.Prayer·J.Pan·C.Herold·H.Prosch
1Department of Biomedical Imaging and Image-Guided Therap y, Computational Imaging Research Lab,
Medical University of Vienna, Vienna, Austria
Machine learning: from radiomics
to discovery and routine
Machine learning is a rapidly
evolving research field attracting
increasing attention in the medical
imaging community. Machine
learning in radiology aims at training
computers to recognize patterns
in medical images and to support
diagnosis bylinking these patterns to
clinical parameters such as treatment
or outcome. These methods enable
the quantification of disease extent
and the prediction of disease course
with higher precision than is possible
with the human eye.
The emergence of machine
learning in radiology
Tworecentadvanceshavefurther acceler-
ated the development of machine learn-
ing in radiology. First, the acquisition
volume of medical imaging data is accel-
erating. Worldwide, during 2000–2007,
an estimated 3.6 billion radiologic, den-
tal radiographic, and nuclear medicine
examinations were performed per year
[18]. Medical imaging data are expected
to soon amount to 30% of worldwide
data storage [9]. Second, recent algorith-
micdevelopmentin the machine learning
field together with new hardware, such
as powerful graphics processing units
(GPU), have yielded a dram atic improve-
mentinthe capabilityof these techniques.
Here, we review the current state of the
art and the possible roles machine learn-
ing can play in medical imaging covering
clinical routine and research. Aer sum-
marizing the basics of machine learning,
we discuss the most pressing challenges
and structure the review around four core
questions:
1. How can machine learning serve as
a tool to perform automated quanti-
tative measurements in radiological
routine?
2. Can machine learning contribute to
research by expanding the vocabu-
lary of patterns we can exploit for
diagnosis and prognosis?
3. Can we effectively expand the evi-
dence on which machine learning
relies from controlled studies to large-
scale routine imaging data?
4. What are the roles of machine
learning in current radiology?
ese questions are connected to a num-
ber of challenges, tackled by research at
the interface of machine learning and ra-
diology. ey range from the availability
of data and expert annotations, to the ex-
ploitation of partially unstructured data
acquire d during routine for learning , and
the coping with noise in both data, as well
as annotations.
Machine learning
In conventional analysis soware, com-
puters are programmed to carry out
instructions specified by the developer.
While machine learning consists of algo-
rithms implemented in standard frame-
works, its approach toward data analysis
is different. Machine le arning algorithms
learn from examples and replicate the
observed behavior—typically a mapping
from an input to an output—on new data.
e basic principles of such models are
illustrated in .Fig. 1.
For instance, training data can consist
of example magnetic resonance imaging
(MRI) volumesdepicting segments of the
breast. For each example, there is a label
indicating whether there is a lesion or
not in the volume. e algorithm learns
a mapping from the input (volume) to
the output (label).
isparadigmiscalledsupervised
learning since a large number of expert
annotations in the form of correct la-
bels are necessary. e resulting model
is called a classifier since it classifies
the input into a discrete set of possible
classes (lesion, no lesion). Aer train-
ing, the classifier can process new input
data and produce an output in the same
categories as it has learned during the
training phase. e first main compo-
nent of these models is feature extraction
or the mapping of raw input data such
as lesion location or delineation to a fea-
ture representation, typically in the form
of a vector. Features capture relevant
characteristics of the raw input data.
Constructing informative features for
specific classification tasks was long the
focus of research and has only recently
shied to being solved by algorithms.
e second component is the mapping
model, such as a classifier in the case of
Abbreviations
AI Artificial intelligence. Machines
that perform tasks typically
requiring human intelligence,
such as planning or problem
solving
DL Deep learning. An approach
for performing ML, where deep
neural network architec tures are
employed
ML Machine learning. An approach
for achieving AI, where instead
of programming task execution
by hand, the machine learns from
examples
Der Radiologe

Review
Fig. 1 9Princ iples of ma-
chine learning
categorical output (benign vs. malign),
or regressionmode ls (time to recurrence)
in the case of continuous output.
By contrast, unsupervised learning
does not rely on expert labels for the
training examples. Instead, it processes
a large number of unlabeled examples
and seeks to find structure in the data.
It can be in the form of clusters, groups
of examples that are similar and clearly
separable from other groups. is cluster
model can then be used to assign cluster
memberships to new data. Bone texture
patterns that can be identified repeatedly
across a population are such an example
[22]. A different result of unsupervised
learning is relationship patterns, oen in
the form of manifolds that reflect the fact
that oen d ifferent variables do not oc cur
in arbitrary combinations; for instance,
age, weight, and height in children.
From feature construction to deep
learning
A range of algorithms for classification
and regression have been investigated
over time, all of them sharing the aim of
learning a mapping from a complex input
space to a label or scalar variable. Near-
est neighbor approaches predict labels
based on the distance of the feature rep-
resentation to those representations with
a known label. Support vector machines
[8]haveprovenimmenselypowerfulin
solving a variety of classification prob-
lems. ey define the decision boundary
in the feature space based on a small
number of support vectors.
e insight that large heterogeneous
training data are oen beyond the ca-
pacity of single classifiers led to two ap-
proaches relying on ensembles of clas-
sifiers: bagging and boosting.Bothuse
a large number of relatively simple mod-
els, so-called weak learners trained on
parts of the data, and aggregate their
estimates on new data, by mechanisms
such as voting. During training, boost-
ing builds a cascade of weak learners,
each trained on data that previous learn-
ers were performing poorly on. An ex-
ample is AdaBoost [7]. Bagging draws
training examples and subsets of features
randomly when training the ensemble of
weak learners. A prominent example is
random forests [3]. e latter introduced
a reliable capability to estimate the in-
formation value of individual predictor
variables for accurate classification. is
led to a wide uptake in communities that
mine large data for predictors, such as
in genomics research, and to a shi in
research efforts from hand-craing fea-
tures toward the exploration of large fea-
ture candidate pools and the algorithmic
selection of predictive features by bag-
ging.
e critical contribution of random
forests was the ability of the algorithm to
work with very large numbers of features,
even if a portion of them is not informa-
tive, and to identify those that carry in-
formation during training. Algorithms
that identify or construct relevant fea-
tures based on examples were shown to
outperform approaches that rely on the
careful design of descriptors capturing
relevant image information.
Most recently, deep learning has
emerged as a powerful approach in ma-
chine learning. It uses neural networks
with large numbers of layers to perform
both feature construction and predictor
Der Radiologe

learning simultaneously [16]. Convo-
lutional neural networks were a first
instance of such architectures used for
supervised learning [15]. Unsupervised
approaches include autoencoders (AE),
a type of neural network that captures
structure by learning to reproduce input
such as images, through a bottle-neck
low-dimensional representation [23],
and more recently generative adversarial
networks (GAN; [6]). With its expand-
ing capabilities, machine learning is an
increasingly important part of research
at the interface of computer science and
medicine (.Figs. 2and 3).
Roles of machine learning in
medical imaging
e computational analysis of imaging
data using machine learning is about to
fundamentally change our ability to un-
derstand disease, risk, and treatment. It
will likely serve as the basis for the de-
velopment and implementation of novel
treatment strategies as well as the indi-
vidualized, effective, and early delivery
of care and preventive medicine. Link-
ing research in machine learning and
medicine will act as a powerful driver
in this endeavor. Interdisciplinary col-
laboration is the key to creating better
models to predict disease course and risk
in individual patients, to forecast the re-
sponse to a growing landscape of possible
treatments, to deliver early prevention,
and to make possible the optimal care
for each individual. We will discuss two
important roles of machine learning in
radiology:
1. e automation of repetitive tasks,
the enabling of radiomics, and the
evaluation of complex patterns in
imaging data not interpretable with
the naked eye.
2. e discovery of new marker patterns
and disease signatures in combined
imaging and clinical record data,
and the linking of these signatures
to disease course and prediction of
treatment response.
Abstract · Zusammenfassung
Radiologe https://doi.org/10.1007/s00117-018-0407-3
© The Author(s) 2018
G. Langs · S. Röhrich · J. Hofmanninger · F. Prayer · J. Pan · C. Herold · H. Prosch
Machine learning: from radiomics to discovery and routine
Abstract
Machine learning is rapidly gaining
importance in radiology. It allows for the
exploitation of patterns in imaging data
and in patient records for a more accurate
and precise quantification, diagnosis, and
prognosis. Here, we outline the basics of
machine learning relevant for radiology,
and review the current state of the art, the
limitations, and the challenges fac ed as these
techniques become an important building
block of precision medic ine. Furthermore, we
discuss the roles machine learning can play in
clinical routine and research and predict how
it might change the field of radiology.
Keywords
Decision support · A rtificial intelligence ·
Computed tomography · Imaging ·
Informatics
Maschinelles Lernen: von Radiomics zu Forschung und Routine
Zusammenfassung
Maschinelles Lernen gewinnt in der
Radiologie rasch an Bedeutung. Es ermöglicht
die Auswertung von Mustern in Daten
aus bildgebenden Untersuchungen und
Patientenakten für eine genauere und
präzisere Quantifizierung, Diagnose und
Prognose. Im vorliegenden Beitrag werden
die für die Radiologie relevanten Grundlagen
des maschinellen Lernens dargelegt und
ein Überblick über den derzeitigen Stand
der Wissenschaft, die Limitationen und die
anstehenden Herausforderungen gegeben,
da sich diese Technik zu einem wichtigen
Baustein der Präzisionsmedizin entwickelt. Es
wird erörtert, welche Rolle das maschinelle
Lernen im Rahmen der klinischen Routine
und in der Forschung spielen kann; darüber
hinaus wird ein Ausblick darauf gegeben,
wie diese Technik die Radiologie langfristig
beeinflussen könnte.
Schlüsselwörter
Entscheidungsunterstützung ·
Künstliche Intelligenz · Computer-
tomographie · Bildgebung · Informatik
Machine learning for
computational quantification
during routine
Machine learning can automaterepetitive
tasks, such as lesion detection, and the
quantific ation of patterns s uch as texture s
that are hard to discern reliably with the
human e ye.
Automation of repetitive tasks
Oneexampleofatedioustask is the detec-
tion, counting, and measurement of le-
sions in large-field-of-view imaging data
before an d aer chem otherapy to e valuate
response. Accuracy has a high impact on
theprognosisandfurther treatmentof the
patient. Nevertheless, the task itself does
not require a high level of experience but
rather a high level of concentration. Ad-
ditionally, finding a lesion and classifying
it as responsive ornot responsive to treat-
ment is not sufficient, as a quantitative
measurement may provide a more accu-
rate description of the patient’s condition.
ere are already several soware pack-
ages available for clinical implementation
[20]. Using this soware, the radiologist
is able to supervise the results, include
a complete quantitative summary in the
radiology report, and focus on interpret-
ing the findings in correlation with the
clinical information. Typically, these ap-
proaches are based on supervised learn-
ing and rely on large manually annotated
training corpora. Recently, approaches
to exploit clinical routine imaging and
record information to train such detec-
tors have shown promising results. ey
rely on weak ly supervis ed learning to lin k
information in images and radiology re-
ports [10].
Der Radiologe

Review
Fig. 2 8Supervise d and unsupervised machine learning
Fig. 3 8Number of pub lications for artificial intelligence, machine learning, and deep learning ac-
cordingto PubMed search. Values for 2018 are extrapolatednumbers based on publications fromJan-
uary to April 2018
Using radiomics features
to quantify subtle image
characteristics
Radiomics, or the high-throughput ex-
traction and analysis (or mining) of ra-
diological imaging data [5], is a step
beyond automating what can be done
with the naked eye. In this process, hun-
dreds of imaging features are calculated
mathematically. ey representinforma-
tion that is not intuitively recognizable.
ereby, a large amount of quantitative
information that was previously not ac-
cessible for human interpretation can be
exploited. e resulting imaging features
belong to specific groups (e.g., shape-
andsize-basedortexturalandwavelet
features, as well as first-order statistics).
While the choice of features to extract
and the subsequent conventional statisti-
cal analysis may be done manually, a ma-
chine learning approach greatly improves
the workflow by automatically extracting
and selecting appropriate and stable fea-
tures. Ultimately it can create predictive
models that rely on features extracted
from imaging data and corresponding
clinical records. At this time, the main
field of application of radiomics is on-
cology and the capture of fine-grained
characteristics of lesions, leading to in-
sights such as a correlation of radiomics
features and the gene expression type of
the imaged lung cancer [1]. However,
this approach has the potential to im-
prove the evaluation and computational
assessment of other diseases too. Exam-
ples are chronic obstructive pulmonary
disease [19] and osteoporosis [21].
Machine learning as a tool for
discovery
In addition to quantifying known pat-
terns, machine learning enables the dis-
covery of new patterns, which can serve
as candidate markers for diagnosis, out-
come prediction, or risk assessment.
Evaluation of complex patterns
and the discovery of new marker
patterns
e diagnosis of disease patterns in radi-
ology relies on the identification of imag-
ing patterns that have proven to be of
diagnostic and/or prognostic relevance.
One prerequisite of such an approach is
thatthepatternsmustbe defined precisely
enough for trained radiologists to be able
to recognize them with low interobserver
variability. e reliable recognition of
disease patterns in radiology is a chal-
lenging task and requires the systematic
analysis of images. A good example of
this is the diagnosis of the usual inter-
stitial pneumonia (UIP) pattern, which
is one of the most important patterns in
diffuse parenchymal lung diseases. e
diagnosis of a UIP pattern relies on the
presence of reticular abnormalities and
honeycombing in a basal and peripheral
predominance [17]. In addition, there
should not be any features more consis-
tent with an alternative diagnosis. Al-
though the definition of a computed to-
mography (CT) pattern of UIP appears
to be straightforward, a number of pub-
lications showed that the interobserver
variability of radiologists is moderate at
best [24,25]
Machine learning enables us to utilize
imaging information that is not recog-
nizable for the human eye and thus new
disease patterns and predictive markers
can be discovered. In lung imaging, ma-
chine learning is a promising approach to
support the CT diagnosis of a UIP pattern
andotherpatternsofdiffuseparenchymal
lung diseases and to predict the course
of these diseases [2]. e potential was
elegantly shown in idiopathic pulmonary
fibrosis, where the increasing pulmonary
vessel volume represents a stronger CT
predictorof functional deteriorationthan
traditional imaging patterns [13]. e
Der Radiologe

strength of machine learning is that the re
is no constraint to described patterns.
Instead, machine learning can identify
patterns that are recognizable with high
reliability, w hich could then serve as a ba-
sis for the diagnosis and prognosis.
Bone diseases are one example, where
bone density is linked to outcome such
as fracture risk but does not capture the
rich variety of trabecular architecture
linked to different diseases. Unsuper-
vised learning is able to identify patterns
that can be extracted repe atedlyac rossin-
dividuals and even scanners [22]. Similar
to image patterns, we can learn struc-
ture in the patient population. Here, at
the patient level, unsupervised machine
learningcanidentifyclustersinchest
CT imaging data linked to radiological
findings, and even clinical parameters
[11].
The role of routine data in
machine learning
e amount of available medical data
has been growing exponentially over the
past few decades, with approximately
118 mil lion C T or M RI exami nations p er
year in the EU [12]. We routinely acquire
comprehensive information character-
izing patients including results from
multimodal imaging data; laboratory,
physiological, and pathological results;
and demographic data. Yet, we have
only limited understanding of how to
combine this complex information from
multiple sources and link it to individual
treatment responseand patient outcome.
Currently, only a small fraction of avail-
able data are used to collect evidence for
supporting personalized medicine or to
enhance clinical decision-making. e
potential associated with these data is
not yet fully reached.
Clinical routine data are a particularly
valuable source of real-life patient his-
tories and connected imaging data cap-
turing disease course and treatment re-
sponse across a wide variety of individual
disease paths. Using routine data is chal-
lenging, since they are not acquired in
a systematic fashion as during clinical
trials, but instead with the sole purpose
of individual patient care. Furthermore,
parts of these data are unstructured, ren-
deringtheiruseformodelingchalleng-
ing. Currently, only a fraction of rou-
tine imaging information is exploited for
research and the development of mod-
els. While there has been a tremendous
advance in image acquisition technol-
ogy during the past five decades, the in-
terpretation of imaging data has hardly
changed. Radiologists focus their re-
ports and summarize individual find-
ings based on specific questions provided
by referring clinicians (e. g., suspected
pulmonary embolism). e information
content of imaging modalities such as
three-dimensional volumes with submil-
limeter slice thickness by far exceeds the
capacities of visual assessment by hu-
mans. ecomplexmultivariaterelation-
ship between clinical information (lab-
oratory results, pathology results, clini-
cal history) and imaging information is
not sufficiently reflected in current stud-
ies. Large parts of multivariate infor-
mation that is critical for personalized
medicine are discarded, restricting evi-
dence to small study populations rather
than exploiting real clinical populations
that are currently largely underexploited.
ere are two reasons for this: on the one
hand the lack of technology to use hetero-
geneous and partly unstructured routine
data for machine learning, and on the
other hand the stumbling blocks of data
management. Machine learning on this
scale makes the integration of partially
unstructured clinical information from
patient records and imaging data neces-
sary [4,14], but to date technology for
this is lacking. Both challenges are cur-
rently being tackled, and we can expect
new kinds of evidence and robustness of
prediction models once we are able to
perform machine learning on this body
of observations.
The future
e increasing use of computer-assisted
diagnosis sowarein radiology raises fear
that such soware might eventually take
over the job of radiologists. ere is no
doubt that 5, 10, and 20 years from now,
the job of radiologists will be very dif-
ferent. In a rapidly evolving specialty
such as in radiology, this is to be ex-
pected. In fact, 20 years ago, the task of
radiologists was different from what it is
today and rapid technical developments
are a hallmark of this field. Yet, the prime
impact of machine learning will not be
in replacing, but rather in enabling radi-
ologists to use more powerful prediction
models, and to rely on more comprehen-
sive patterns that are informative for the
individual disease course and treatment
outcome. e radiologist’s role will be to
develop such diagnostic paradigms, and
to integrate these paradigms with overall
patient care.
e primary role of machine learning
will not be limited to the automation of
specific analysis tasks, but will be critical
in improving individual care by discov-
ering new marker patterns and render-
ing these patterns useful. We will have
powerful computational models that will
enable a more accurate prediction of in-
dividual disease course, and which will
forecast a response to treatment from
observations such as imaging data and
diseasehistory. esemodelswillfacil-
itate the selection of personalized treat-
ment strategies more effectively based on
an assessment of each patient in light of
an unprecedented scale of evidence from
thousands of disease and treatment his-
tories, together with models that trans-
late these observations into an accurate
prognosis.
As we become able to detect and quan-
tify complex but subtle patterns in ob-
servations at an earlier stage, diagnostic
categories might change. rough an
understanding of the link between dis-
tributed signatures andprognosis, we will
discover novel disease and response phe-
notypes, which, in turn, will contribute
to research for novel treatments.
Until then, a number of hurdles must
be overcome. Research communities that
advance machine learning and medical
disciplines such as radiology must col-
laborate more closely, learning to trig-
ger algorithmic methodology by clinical
questions and to ask new clinical ques-
tions whose answers will be made pos-
sible by computational algorithms. We
will need to find methods to deal with
heterogeneous multimodal data and the
means for the efficient and possibly auto-
mated curation of such data. Algorithms
will have to identify prognostic feature
Der Radiologe

Review
patterns in complex imaging data, and
robust models will need to be able to pre-
dict, simulate, and assess the certainty of
theiroutputatthe same time. Whilethese
challenges are s ignificant, an inc reasingly
joint effort by researchers across varying
fields makes the solution to such chal-
lengesplausible, and thefuture of medical
imaging might change more rapidly than
we expect today.
Practical conclusion
4Machine learning facilitates the
exploitation of patterns in imaging
data and patient records for more
accurate andprecise quantification,
diagnosis, and prognosis.
4The computational analysis of imag-
ing data using machine learning will
change our ability to understand
disease, risk, and treatment.
4Machine learning allows for the
automation of repetitive tasks, the
enabling of radiomics, and the
evaluation of complex patterns
in imaging data not interpretable
with the naked eye. It leads to the
discovery of new marker patterns
and disease signatures in imaging
and clinical record data, and the
linking of these signatures to disease
course and prediction of treatment
response.
4Radiologists will be able to use
more powerful prediction models
and to rely on more comprehensive
patterns providing information on
the individual disease course and
treatment outcome.
Corresponding address
Prof. Dr. G. Langs
Department of Biomedical Imaging and Image-
Guided Therapy, Computational Imaging
Research Lab, Medical Universi ty of Vienna
Währinger Gürtel 18–20, 1090 Vienna, Austria
georg.langs@meduniwien.ac.at
Funding . Open access funding provided by Medical
University of Vienna. Part of this work was funded by
the Austrian Science Fund FWF(I 2714-B31).
Compliance with ethical
guidelines
Conflict of interest. G. Langs, S. Röhrich, J. Hofman-
ninger,F. Prayer,J. Pan, C. Herold, and H.Prosch declare
that they have no competinginterests.
This article does notcontain any studies with human
participants or animals performed by any of the au-
thors.
The supplementcontaining this article is not spon-
sored by industry.
Open Access.This article is distributedunder the terms
of the Creative Commons Attribution 4.0 International
License (http://creativecommons.org/licenses/by/
4.0/), which permitsunrestricted use, distribution,
and reproduction in any medium, provided you give
appropriate creditto the original author(s) and the
source,provide a link to theCreative Commons license,
and indicate if changes were made.
References
1. Aerts HJWL et al (2014) Decoding tumour pheno-
type by noninvasive imaging using a quantitative
radiomics approach. NatCommun 5:4006
2. Bartholmai BJ, Raghunath S, Karwoski RA et al
(2013) Quantitative computed tomography
imaging of interstitial lung diseases. J Thorac
Imaging 28(5):298–307. https://doi.org/10.1097/
RTI.0b013e3182a21969
3. Breiman L (2001) Random forests. Mach Learn
45(1):5–32
4. European Society of Radiology (2011) Medical
imaging in personalised medicine: a white paper
ofthe researchcommitteeofthe EuropeanSociety
ofRadiology(ESR). Insights Imaging 2(6):621–630
5. Gillies RJ, Kinahan PE, Hricak H (2016) Radiomics:
images are more than pictures, they are data.
Radiology 278(2):563–577
6. Goodfellow I et al (2014) Generative adversarial
nets. In: Ghahramani Z et al (ed) Advances in
neural informationprocessing systems 27. Curran
Associates,Inc, Morehouse,pp2672–2680
7. Hastie T et al (2009) Multi-class AdaBoost. Stat
Interface 2(3):349–360
8. Hearst MA et al (1998) Support vector machines.
IEEEIntell Syst Appl 13(4):18–28
9. High-Level Expert Group on Scientific Data
(2010) Riding the wave—how europe can gain
from the rising tide of scientific data. Final
Report to the European Commission. p. 1–40.
https://www.researchgate.net/publication/
255181186_Riding_the_wave_How_Europe_
can_gain_from_the_rising_tide_of_scientific_
data_Final_report_of_the_High_Level_Expert_
Group_on_Scientific_Data_A_submission_to_
the_European_Commission. Accessed: 12 June
2018
10. Hofmanninger J et al (2015) Mapping visual
featuresto semanticprofiles forretrievalin medical
imaging. Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition,
pp457–465
11. Hofmanninger J et al (2016) Unsuper vised iden-
tification of clinically relevant clusters in routine
imaging data. In: Medical Image Computing and
Computer-Assisted Intervention—MICCAI 2016.
Springer,Cham, pp192–200
12. https://data.oecd.org/healthcare/computed-
tomography-ct- exams.htm. Accessed: 12 June
2018
13. Jacob J, Bartholmai BJ, Rajagopalan S et al
(2018) Serial automated quantitative CT analysis
in idiopathic pulmonary fibrosis: functional
correlations and comparison with changes in
visual CT scores. Eur Radiol 28(3):1318–1327.
https://doi.org/10.1007/s00330-017-5053-z
14. Knechtges PM, Carlos RC (2007)The evolving role
of radiologists within the health care system. J Am
CollRadiol 4(9):626–635
15. Krizhevsky A, Sutskever I, Hinton GE (2012)
Imagenet classification with deep convolutional
neural networks. In: Pereira F et al (ed) Advances in
neural informationprocessing systems 25. Curran
Associates,Inc, Morehouse,pp1097–1105
16. LeCun Y, Bengio Y, Hinton G (2015)Deep learning.
Nature521:436
17. Lynch DA, Sverzellati N, Travis WD et al (2018)
Diagnostic criteria for idiopathic pulmonary
fibrosis: a Fleischner Society White Paper. Lancet
RespirMed 6(2):138–153
18. Mettler FA Jr et al (2009) Radiologic and nu-
clear medicine studies in the United States and
worldwide: frequency, radiation dose, andcom-
parisonwith otherradiationsources—1950–2007.
Radiology253(2):520–531
19. Ostridge K et al (2016) Relationship between
pulmonary matrix metalloproteinases and quan-
titative CT markers of small airways disease and
emphysemain COPD.Thorax71(2):126–132
20. Rubin DL et al (2014) Automated tracking of
quantitative assessments of tumor burden in
clinicaltrials. Transl Oncol7(1):23–35
21. Takasu M et al (2014) Multidetectorcomputed to-
mography-based microstructuralanalysis reveals
reduced bone mineral content and trabecular
bone changes in the lumbar spine after transarte-
rialchemoembolization therapy for hepatocellular
carcinoma.PLoS ONE 9(10):e110106
22. Valentinitsch A et al (2013) Computational
identification and quantification of trabecular
microarchitectureclasses by 3-D texture analysis-
basedclustering. Bone 54(1):133–140
23. Vincent P et al (2010) Stacked denoising autoen-
coders: learning useful representations in a deep
network with a local denoising criterion. J Mach
LearnRes 11(Dec):3371–3408
24. Walsh SL, Calandriello L, Sverzellati N, Wells
AU, Hansell DM; UIP Observer Consort (2016)
Interobserver agreement for the ATS/ERS/JRS/
ALAT criteria for a UIP pattern on CT. Thorax
71(1):45–51. https://doi.org/10.1136/thoraxjnl-
2015-207252
25. Watadani T, Sakai F, Johkoh T et al (2013)
Interobserver variability in the CT assessment
of honeycombing in the lungs. Radiology
266(3):936–944. https://doi.org/10.1148/radiol.
12112516
Der Radiologe