INT J TUBERC LUNG DIS 22(3):328–335

Q 2018 The Union

http://dx.doi.org/10.5588/ijtld.17.0520

Detection of tuberculosis patterns in digital photographs of

chest X-ray images using Deep Learning: feasibility study

A. S. Becker,* C. Blüthgen,* V. D. Phi van,* C. Sekaggya-Wiltshire,† B. Castelnuovo,† A. Kambugu,†

J. Fehr,‡ T. Frauenfelder*
*Institute of Diagnostic and Interventional Radiology, University Hospital of Zurich, Zurich, Switzerland;
†
Infectious Disease Institute, College of Health Sciences, Makerere University, Kampala, Uganda; ‡Division of
Infectious Diseases and Hospital Epidemiology, University Hospital Zurich, University of Zurich, Zurich, Switzerland

SUMMARY

O B J E C T I V E : To evaluate the feasibility of Deep Learn-
ing-based detection and classification of pathological
patterns in a set of digital photographs of chest X-ray
(CXR) images of tuberculosis (TB) patients.
M A T E R I A L S A N D M E T H O D S : In this prospective,
observational study, patients with previously diagnosed
TB were enrolled. Photographs of their CXRs were
taken using a consumer-grade digital still camera. The
images were stratified by pathological patterns into
classes: cavity, consolidation, effusion, interstitial chang-
es, miliary pattern or normal examination. Image
analysis was performed with commercially available
Deep Learning software in two steps. Pathological areas
were first localised; detected areas were then classified.
Detection was assessed using receiver operating charac-
teristics (ROC) analysis, and classification using a
confusion matrix.
R E S U LT S : The study cohort was 138 patients with
human immunodeficiency virus (HIV) and TB co-
infection (median age 34 years, IQR 28–40); 54 patients
were female. Localisation of pathological areas was
excellent (area under the ROC curve 0.82). The software
could perfectly distinguish pleural effusions from intra-
parenchymal changes. The most frequent misclassifica-
tions were consolidations as cavitations, and miliary
patterns as interstitial patterns (and vice versa).
C O N C L U S I O N : Deep Learning analysis of CXR photo-
graphs is a promising tool. Further efforts are needed to
build larger, high-quality data sets to achieve better
diagnostic performance.
K E Y W O R D S : CXR; teleradiology; TB; Deep Learning;
chest radiograph

IT IS ESTIMATED THAT, WORLDWIDE, 8 million
people develop active tuberculosis (TB) every year.
The disease, if left untreated, has a mortality rate of
~70%,1 with 2 million people dying each year from
TB or its complications. As TB is only contagious in
its active, pulmonary form, chest X-ray (CXR)
images play a crucial role in identifying affected
patients and taking preventive measures.2,3
The prevalence of TB is highly dependent on socio-
economic factors. This is evident by its decline in
Western Europe in the past century with increasing
wealth, long before effective chemotherapies were
available.4 Today, the majority of the disease burden
is concentrated in only six countries,5 one of the
common denominators being large areas with poor
socio-economic development. In these areas, very few
health care facilities and physicians are available.
However, with the advent of smartphones and mobile
internet, it may be feasible to offer help in the form of
automated diagnosis using the integrated camera of
these handheld devices (Figure 1). As the sending and
receiving of images between mobile phones and a
server can be implemented easily today, we focus on
the actual image analysis (Figure 1, right-hand
image).
The human visual system allows for a very high
level of abstraction.6 Radiologists can therefore
detect abnormalities on original X-ray examinations
as well as on digital photographs of X-ray films.7 As
an example, to the reader of this article the two CXR
images in the top row of Figure 2 will look very
similar. To the computer, however, these images are
fundamentally different, as indicated by the histo-
grams in the middle row. Not only does the digital
photograph come at a lower resolution than the
original image, the addition of spurious colour
channels by the digital camera adds unnecessary
noise and lowers the signal-to-noise ratio (SNR),
which may lower the ability of both computers and
humans to detect subtle lesions close to the inherent

Correspondence to: Anton S Becker, Institute for Diagnostic and Interventional Radiology, University Hospital Zurich,
Raemistrasse 100, 8091 Zurich, Switzerland. e-mail: anton.becker@usz.ch
Article submitted 26 July 2017. Final version accepted 15 November 2017.
 Deep Learning on TB CXR photographs 329

Figure 1 Proposed image processing pipeline. Possible workflow for a teleradiological TB

detection service for health care providers working remotely (schematic). After taking a
photograph of the original chest X-ray with a smartphone application, the image is transferred
wirelessly to a remote server, where image analysis is carried out in two steps: anomaly detection
and classification. The server then sends the image with an overlaid heatmap and a structured
report back to the user (health care provider), who may use it to augment his/her clinical
judgement for optimal management of the patient.

noise limit (bottom row). Furthermore, other objects
in the image (such as the Post-itw Notes in our
example) may add further ‘real-world noise’. For a
computer algorithm to be successful in this task, it
would thus need to exhibit a high ability for
abstraction, similar to that for human readers. A
novel machine-learning technique called Deep Learn-
ing has recently yielded remarkable results, with its
performance becoming increasingly human-like in
recent years.8
The purpose of the present study was to evaluate
the feasibility of Deep Learning-based detection and
classification of pathological patterns in a set of
digital photographs of CXR images of TB patients
from Uganda.
STUDY POPULATION AND METHODS
Patient population
This was a prospective observational study of patients
enrolled at the integrated TB-HIV (human immunode-
ficiency virus) out-patient clinic of the Infectious
Diseases Institute in Kampala, Uganda. CXRs were
taken in the Radiology Department of Mulago
National Referral Hospital, which is the main national
and teaching hospital providing specialised care to the
whole country. The study received ethics approval from
the Joint Clinical and Research Centre Ethics Com-
mittee, Kampala, and the Uganda National Council for
Science and Technology, Kampala (HS 1303).
All photographs were taken using a digital still
camera (DMC-TZ56; Panasonic, Osaka, Japan; see
Table 1 for full details). The images were screened by
two investigators in consensus (ASB, TF) and
stratified by the dominant pathological pattern:
cavity, consolidation, effusion, interstitial changes,
miliary pattern or normal examination.
Image analysis
For image analysis, industrial-grade Deep Learning
image analysis software (Suite v2.0; ViDi Systems,
Villaz-Saint-Pierre, Switzerland) was used.9,10 In Deep
Learning, artificial neural networks are arranged in
multiple layers, which imitates the brain, their natural
counterpart.11 The neurons of the mammalian neocor-
tex are organised in multiple layers, which is particu-
larly apparent in the human visual cortex. This means
that any given input percept will be deconstructed and
represented at multiple levels of abstraction, each
corresponding to a different layer or cortical area.6
Although currently not approved for routine clinical
use, ViDi has recently shown human-like performance
in the detection of breast cancer.12
Computations were carried out using a GeForce
GTX 1080 graphics processor unit (Nvidia, Santa
Clara, CA, USA). Anomalies were first contoured by
two investigators in consensus (ASB, TF) using the
ViDi Detection tool for supervised training. A
randomly chosen subset of images (n ¼ 117, 85%)
was used to train the software, and the remaining
cases (n ¼ 21) were used to validate the resulting
model (cross-validation). The detected anomalies
were then transferred as individual image patches to
the ViDi classification tool and labelled by the same
investigators. Again, a randomly chosen subset of
image patches (n ¼ 150, 80%) was used to train the
software, and the remaining patches (n ¼ 40) were
used for validation.
330 The International Journal of Tuberculosis and Lung Disease

Figure 2 Illustrated comparison of original X-ray and photograph. Examples of an original digital chest X-ray image (DCXr, top left)
and a digital photograph of a film radiograph (photo, top right). Although to the human observer the images look very much alike, the
histogram (middle row) indicates that they look fundamentally different to a computer. While the DCXr only contains greyscale
information, a photograph typically contains three additional separate colour channels, which are an ‘artifact’ of the camera sensor
and do not convey any meaningful information, as the original radiograph only contains greyscale information as well. This scenario
results in a decreased signal-to-noise ratio (bottom row) and hence a smaller real-world information content per image, which lowers
the conspicuousness of subtle differences/lesions (top right and bottom left circles).

Table 1 Technical details of the Panasonic Lumix DM-TZ57 camera*

Sensor Sensor size/total pixels/filter 1/2.33-type high-sensitivity MOS sensor/total pixel number 17.5 megapixels/primary colour filter
Pixels Camera effective pixels 16 megapixels
Lens
Aperture F3.3–6.4/multistage iris diaphragm (F3.3–8.0 (W), F6.4–8.0 (T))
Optical zoom 203
Focal length f ¼ 4.3–86.0 mm (24–480 mm in 35 mm equivalent)
Lens Lumix DC Vario/12 elements in 10 groups/(3 aspherical lenses/6 aspherical surfaces/2 extra-low
dispersion lenses)
Focus
Focusing area Normal: wide 50 cm–infinity/tele 200 cm–infinity/AF macro/intelligent auto/motion picture: wide 3
cm– infinity/tele 100 cm–infinity
Focus Normal/AF macro/macro zoom/quick AF (always on), continuous AF (only for motion picture)/AF
tracking
Shutter Shutter speed Approximately 4–1/2000 s (0.2 s)
Exposure parameters
ISO sensitivity Auto/i.ISO/100/200/400/800/1600/3200/high-sensitivity mode (ISO1600–6400)
White balance Auto/daylight/cloudy/shade/incandescent/white set/white balance adjustment (except auto)
Flash Not used
* Panasonic, Osaka, Japan.
MOS ¼ metal–oxide–semiconductor; AF ¼ auto focus; ISO ¼ International Organization for Standardization.
 Deep Learning on TB CXR photographs 331

Table 2 Disease patterns in the study cohort Image analysis

Pattern
Cavity
Consolidation
Effusion
Interstitial
Miliary
None (controls)
Statistical analysis
Due to the preliminary nature of the present study, the
statistical analysis is mainly descriptive. Continuous
variables are expressed as mean and standard devia-
tion; categorical variables as counts and percentages.
Detection performance was assessed using receiver
operating characteristics (ROC) analysis with the area
under the curve (AUC). Pathology classification was
assessed with a confusion matrix generated using
Python 3.5.2, seaborn 0.7.1, numpy 1.12.0 and pandas
0.19.2 packages (Python Software Foundation, Beaver-
ton, OR, USA). The specificity and positive predictive
value (PPV) were calculated for each feature in the
whole data set as well as for the validation data only.
RESULTS
Study cohort
Our cohort comprised 138 HIV-infected patients
diagnosed with TB (median age 34 years, interquar-
tile range 28–40). Over 95% of these participants
presented to the clinic with a cough of 72 weeks’
duration, and .80% had fever and weight loss, in
accordance with their diagnosis of TB. The number of
different patterns is given in Table 2.
n The training time for pathology detection was 2 min.
10 The processing time per validation image with the
68 final model was 53.4 6 1.4 msec. Detection
18 performance was excellent, with AUC ¼ 0.98
14
8 (validation set 0.82) (Figure 3A).
20 More than one feature was often present in the
patches detected (e.g., consolidation around a cavity,
as shown in Figure 4A, which looks similar to the
pure consolidation in Figure 4B). As the software
does not support multiple labels per patch, and due to
the limited size of our data set, the most dominant
feature was used when labelling the patches. Training
time was 3 min; processing time per validation image
was 39.4 6 9.7 msec. The confusion matrix is
illustrated in Figure 3B: overall, the software was
able to perfectly distinguish pleural effusions from
intraparenchymal patterns (specificity and PPV ¼ 1.0
overall and in the validation set, see example in Figure
4C). The most frequent false diagnoses were cavita-
tions misclassified as consolidations or vice versa (n ¼
4), and a miliary pattern mistaken for an interstitial
pattern (n ¼ 5) and vice versa, resulting in the
markedly lower sensitivity and PPV for these patterns
(Table 3). Furthermore, miliary and interstitial
patterns were sometimes misclassified as consolida-
tions, but not vice versa (see example in Figure 4D).
DISCUSSION
The detection and classification of pathology as seen
in digital photographs of CXR images of TB patients
is feasible with Deep Learning. Despite the draw-

Figure 3 A) Receiver operating characteristic curve of disease detection in the whole cohort demonstrates the excellent ability of the
software to localise abnormal areas in the chest X-ray photograph (AUC 0.98). B) Confusion matrix of the classification of the
different patterns. Pleural effusions were detected with excellent accuracy (1.0), whereas cavities were sometimes misdiagnosed as
consolidations (probably due to surrounding consolidations). Diagnosis of interstitial and miliary patterns also exhibited a greater
degree of uncertainty. AUC ¼ area under the receiver operating characteristic curve.
332 The International Journal of Tuberculosis and Lung Disease

Figure 4 A) This cavity with surrounding consolidations in the right upper lobe was correctly
detected and diagnosed. B) Sample image of consolidations in the left upper lobe, which were
detected and classified correctly by the software. C) An example of a small left-sided pleural
effusion, which was reliably detected and classified. D) Female patient with extensive bilateral
miliary tuberculosis. Although pathological changes were detected correctly, they were falsely
identified as consolidations.

Table 3 Sensitivity, specificity, PPV and AUC for the different patterns overall (all) and untrained
samples only (valid)
Sensitivity
All Valid
Cavity 0.86 0.50
Consolidation 0.95 0.74
Effusion 1.00 1.00
Interstitial 0.88 0.43
Miliary 0.84 0.29
PPV ¼ positive predictive value; AUC ¼ area under the ROC curve.
backs of digital photography, image analysis yielded
useful preliminary results.
Due to the often negative smear results in HIV/
AIDS (acquired immune-deficiency syndrome) pa-
tients, CXR may be the main tool for establishing a
diagnosis of active TB in these patients and is
endorsed by the World Health Organization,13
despite some conflicting evidence in the literature.14
One possible reason for this could be the shortage of
expert physicians in affected areas. Non-expert
readers, even after several training sessions and
implementation of a systematic reading and reporting
system, may exhibit only limited diagnostic accuracy
in detecting X-ray signs of active TB.15–17 With
rapidly expanding access to smartphone cameras and
mobile internet in the developing world, diagnosing
active TB with the help of these tools appears to be a
logical next step in resource-limited settings.
However, digitising the film using a consumer-
grade digital camera sensor adds spurious colour
channels and electronic noise to the original greyscale
image. In larger centres with an existing digital
workflow and a medical-grade film digitiser, the
‘smartphone approach’ may be of limited use, as
Deep Learning can be directly integrated using the
original Digital Imaging and Communications in
Medicine (DICOM) files over a local area network
(LAN) connection. This has already been practised
for decades in Europe and the United States using
traditional computer-aided detection (CAD) pro-
grammes. Due to the higher SNR, more homoge-
neous image characteristics and faster data transfer,
more reliable results with less training data could be
expected. However, the majority of health care
facilities in the developing world remain dependent
on plain-film CXR examinations and a non-digital
workflow. Studying the challenges associated with an
indirect workflow via a digital (smartphone) camera
thus seems justified.
Machine learning is a field in computer science in
which the computer is trained with sample data
instead of being explicitly programmed to perform a
task. The machine learning algorithm, Deep Learning
in our case, analyses the existing relationships and
features of training data and attempts to find useful
patterns.18 This data-driven approach to problem
Deep Learning on TB CXR photographs 333
Specificity PPV AUC
All Valid All Valid All Valid
0.98 0.92 0.67 0.25 0.98 0.90
0.95 0.76 0.95 0.74 0.95 0.75
1.00 1.00 1.00 1.00 1.00 1.00
0.97 0.85 0.85 0.38 0.95 0.78
0.99 0.94 0.93 0.50 0.96 0.83
solving has been tremendously successful in the past
few years in financial modelling, marketing and
education, as well as manufacturing and medical
imaging.19 The latter two fields have a key compo-
nent in common: there is often only a small amount of
high-quality data available for a given manufacturing
line, or body part and modality. To put things in
perspective, in machine learning research, a large data
set would contain ~14 million annotated images with
750 000 samples per high-level category (‘ImageNet
Summary and Statistics’. http://image-net.org/about-
stats accessed July 2017).
We therefore chose to evaluate Deep Learning
software conceived for industrial use because we
hypothesised that due to these shared problems, our
medical image analysis may benefit from the solu-
tions engineered into such software. The drawback to
this method is the proprietary network architecture,
which makes it impossible to understand or improve
the underlying algorithm itself. In contrast, current
state-of-the-art open-source machine learning frame-
works such as TensorFlow (Google, Mountain View,
CA, USA) or Keras (Massachusetts Institute of
Technology, Cambridge, MA, USA) enable access to
the network architecture and hyperparameters as well
as visualisation of activation patterns in each layer of
the network. We project that such freely available,
open-source software will play an important role in
the research and development of machine learning
tools for the detection of TB in CXR examinations.
There are three inherent limitations to Deep
Learning. First, especially in smaller data sets, there
is a significant danger of overfitting the model. This
means that the trained network will not generalise
well to new, unseen cases. Usually, this is assessed by
(cross-)validation, i.e., setting aside a percentage of
the data ‘not shown’ to the algorithm and evaluating
the performance on this validation set. Second, the
exact computations that take place are, due to the
sheer complexity of the network, not traceable or
comprehensible for the human observer. This means
that the resulting model is something of a ‘black box’.
The use of heat maps which highlight the features
learned to be suspicious by the neural network may
somewhat compensate for this drawback.20 Third,
many authors use excessive downsampling of the
334 The International Journal of Tuberculosis and Lung Disease
images. For example, in a recent study on identifying
TB patterns in CXR images using Deep Learning by
Lakhani and Sundaram, images were scaled down to
a resolution of 256 3 256 pixels.21 This is usually
necessary to improve efficiency in terms of memory
and computing time. However, in contrast to, for
example, object identification in the real world,
diagnostic imaging often relies on the detection of
subtle changes. In particular, initial changes in early
disease stages are often only detectable at full
resolution. It may thus be preferable to build
networks which take advantage of the high resolution
as, for example, in a recent study on mammogra-
phy.22
The first step, the detection of pathological
patterns, worked very well, despite the small number
of cases. Keeping in mind that the neural networks of
the software were engineered to find anomalies in
highly repetitive textures and patterns, this remains
surprising, as CXR examinations are composed of
fairly complex patterns with a high degree of natural
and normal variability. Moreover, performance was
slightly higher (AUC 0.82) than traditional computer-
aided systems specifically developed for TB detection
(AUC 0.71–0.75),23,24 which is comparable with
human performance.25 When classifying the different
manifestations of the disease, our results showed a
clear strength in differentiating between effusion and
parenchymal changes. As demonstrated by the worse
performance in the validation set, differentiation
between intraparenchymal patterns proved to be
more challenging for three main reasons. First, for
our pilot study, only a limited data set was available.
We therefore expect our performance to improve with
more experience (which equates to more data), for
which we have planned a follow-up study. Second,
purely isolated patterns are rarely present in real-
world clinical scenarios. Cavities may often be
surrounded by consolidations or prominent intersti-
tial markings or, in addition to a miliary pattern,
consolidations or thickened interstitial septae may be
present. This leads to a certain degree of ‘confusion’
in the neural network due to the inherent imprecision
of the data. As a result, inexperienced human readers
also exhibit poor consensus with regard to character-
ising intraparenchymal patterns in bedside CXR
images but not when diagnosing pleural effusions.26
Finally, small fluid- or debris-filled cavities could be
present within areas of consolidation. As Deep
Learning inherits the physical limitations of the given
modality, the density difference between fluid and
soft tissue is too small to be discriminated on plain X-
ray examinations.
In our study, we focused our analysis on active
airspace disease and pleural effusions. In a clinical
setting, these are probably the most important
findings, with immediate consequences for patient
management. Future studies could also investigate
mediastinal involvement (enlarged and/or calcified
lymph nodes or pericardial effusions), airway affec-
tion (bronchiectasis or tracheobronchial stenosis) or
sequelae such as aspergilloma (present in up to 11%
of patients with chronic disease27) or scar-based
changes. For the latter category, it would be helpful to
evaluate the course of the disease on the basis of serial
CXR images, a problem suitable for recurrent neural
networks.28
In conclusion, analysis of even a few CXR
photographs from TB patients is feasible using Deep
Learning software. Although further efforts are
needed to build larger, high-quality data sets and
achieve better diagnostic performance, our approach
may be useful to develop a machine learning-based
application to help diagnose TB in remote or
underdeveloped areas.
Acknowledgements
The authors thank all patients for their participation and their
families for their support, as well as the health care workers at
Infectious Diseases Institute, Kampala, Uganda; M Bauer for image
digitalisation and N Eberhard for her support while working in
Uganda.
Conflicts of interest: none declared.
References
1 Tiemersma E W, van der Werf M J, Borgdorff M W, Williams B
G, Nagelkerke N J D. Natural history of tuberculosis: duration
and fatality of untreated pulmonary tuberculosis in HIV
negative patients: a systematic review. PLOS ONE 2011; 6:
e17601.
2 Skoura E, Zumla A, Bomanji J. Imaging in tuberculosis. Int J
Infect Dis 2015; 32: 87–93.
3 Somashekar N, Chadha V K, Praseeja P, et al. Role of pre-
Xpertw screening using chest X-ray in early diagnosis of smear-
negative pulmonary tuberculosis. Int J Tuberc Lung Dis 2014;
18: 1243–1244.
4 Davies P, Gordon S B, Davies G. Clinical tuberculosis. Boca
Raton, FL, USA: CRC Press Book, 2014.
5 World Health Organization. Global tuberculosis report, 2016.
WHO/HTM/TB/2016.13. Geneva, Switzerland: WHO, 2016.
6 Serre T, Kreiman G, Kouh M, Cadieu C, Knoblich U, Poggio T.
A quantitative theory of immediate visual recognition. In: Cisek
P, Drew T, Kalaska J, eds. Computational neuroscience:
theoretical insights into brain function. Progress in Brain
Research. Vol 165. Amsterdam, The Netherlands: Elsevier,
2007: pp 33–56.
7 Szot A, Jacobson F L, Munn S, et al. Diagnostic accuracy of
chest X-rays acquired using a digital camera for low-cost
teleradiology. Int J Med Inform 2004; 73: 65–73.
8 Esteva A, Kuprel B, Novoa R A, et al. Dermatologist-level
classification of skin cancer with deep neural networks. Nature
2017; 542: 115–118.
9 LeCun Y A, Bottou L, Orr G B, Müller K-R. Efficient BackProp.
In: Neural networks: tricks of the trade. Lecture Notes in
Computer Science. Vol 7700. Berlin, Heidelberg, Germany:
Springer Berlin Heidelberg, 2012: pp 9–48.
10 Hinton G E. To recognize shapes, first learn to generate images.
Prog Brain Res 2007; 165: 535–547.
11 Bengio Y. Learning Deep architectures for AI. In: Jordan M, ed.
Foundations and trendsw in machine learning. Delft, The
Netherlands: Now Publishers Inc, 2009.

12 Becker A S, Marcon M, Ghafoor S, Wurnig M C, Frauenfelder
T, Boss A. Deep Learning in mammography: diagnostic
accuracy of a multipurpose image analysis software in the
detection of breast cancer. Invest Radiol 2017; 52: 434–440.
13 Harries A D, Maher D, Graham S. TB/HIV: a clinical manual.
Geneva, Switzerland: World Health Organization, 2004.
14 Davis J L, Worodria W, Kisembo H, et al. Clinical and
radiographic factors do not accurately diagnose smear-negative
tuberculosis in HIV-infected inpatients in Uganda: a cross-
sectional study. PLOS ONE 2010; 5: e9859.
15 Kosack C S, Spijker S, Halton J, et al. Evaluation of a chest
radiograph reading and recording system for tuberculosis in a
HIV-positive cohort. Clin Radiol 2017; 72: 519.e1–519.e9.
16 Semakula-Katende N S, Andronikou S, Lucas S. Digital
platform for improving non-radiologists’ and radiologists’
interpretation of chest radiographs for suspected tuberculosis—
a method for supporting task-shifting in developing countries.
Pediatr Radiol 2016; 46: 1384–1391.
17 Seddon J A, Padayachee T, Plessis du A-M, et al. Teaching chest
X-ray reading for child tuberculosis suspects. Int J Tuberc Lung
Dis 2014; 18: 763–769.
18 Kohli M, Prevedello L M, Filice R W, Geis J R. Implementing
machine learning in radiology practice and research. AJR Am J
Roentgenol 2017; 208: 754–760.
19 Jordan M I, Mitchell T M. Machine learning: trends,
perspectives, and prospects. Science 2015; 349: 255–260.
20 Zhou B, Khosla A, Lapedriza A, Oliva A, Torralba A. Learning
deep features for discriminative localization. 2016 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), 27–30 June 2016, Las Vegas, NV, USA: pp 2921–
2929.
Deep Learning on TB CXR photographs 335
21 Lakhani P, Sundaram B. Deep Learning at chest radiography:
automated classification of pulmonary tuberculosis by using
convolutional neural networks. Radiology 2017; 284: 574–
582.
22 Geras K J, Wolfson S, Kim S G, Moy L, Cho K. High-resolution
breast cancer screening with multi-view deep convolutional
neural networks. Computer Vision and Pattern Recognition, 6
November 2017.
23 Steiner A, Mangu C, van den Hombergh J, et al. Screening for
pulmonary tuberculosis in a Tanzanian prison and computer-
aided interpretation of chest X-rays. Public Health Action
2015; 5: 249–254.
24 Muyoyeta M, Maduskar P, Moyo M, et al. The sensitivity and
specificity of using a computer aided diagnosis program for
automatically scoring chest X-rays of presumptive TB patients
compared with Xpert MTB/RIF in Lusaka Zambia. PLOS ONE
2014; 9: e93757.
25 Maduskar P, Muyoyeta M, Ayles H, Hogeweg L, Peters-Bax L,
van Ginneken B. Detection of tuberculosis using digital chest
radiography: automated reading vs. interpretation by clinical
officers. Int J Tuberc Lung Dis 2013; 17: 1613–1620.
26 Martini K, Ganter C, Maggiorini M, et al. Interpretation of
bedside chest X-rays in the ICU: is the radiologist still needed? J
Clin Imaging 2015; 39: 1018–1023.
27 Fraser R S, Muller N L, Colman N, Pare P D. Fraser and Paré’s
diagnosis of diseases of the chest. Vols 1–4. Philadelphia, PA,
USA: W B Saunders, 1999.
28 Schmidhuber J. Deep Learning in neural networks: an overview.
Neural Networks 2015; 61: 85–117.

OBJECTIF : Evaluer la faisabilité de l’apprentissage de
la détection et de la classification de profils
pathologiques sur un ensemble de photographies
numériques de radiographie pulmonaires (CXR) de
patients atteints de tuberculose (TB).
M A T E R I E L S E T M E T H O D E S : Dans cette étude
prospective d’observation, les patients atteints de TB
déjà diagnostiquée ont été enrôlés. Les photographies de
leurs CXR ont été prises avec un appareil photo
numérique grand public. Les images ont été stratifiées
en fonction des profils pathologiques dans les classes
suivantes : caverne, consolidation, épanchement,
modifications interstitielles, miliaire ou aspect normal.
L’analyse des images a été réalisée avec un logiciel
d’apprentissage disponible dans le commerce en deux
étapes. D’abord, les zones pathologiques ont été
localisées, puis les zones détectées ont été classées. La
détection a été évaluée par une analyse de fonction
O B J E T I V O: Evaluar la viabilidad de un método basado
en el aprendizaje profundo, con el fin de detectar y
clasificar de perfiles patológicos en un conjunto de
fotografı́as digitales de radiografı́as de tórax (CXR) de
pacientes con tuberculosis (TB).
M A T E R I A L E S Y M É T O D O S: En el presente estudio
prospectivo observacional, se incluyeron pacientes con
un diagnóstico anterior de TB. Se tomaron fotografı́as
de las CXR de los pacientes con una cámara fotográfica
digital de calidad para el consumidor. Las imágenes se
estratificaron por perfiles patológicos en las siguientes
clases: caverna, consolidación, derrame, infiltrados
intersticiales, patrón miliar o examen normal. Se llevó
a cabo un análisis de las imágenes con un programa
informático de aprendizaje profundo en dos etapas. En
primer lugar, se localizaron las zonas patológicas y luego
se clasificaron. Se evaluó la detección mediante una
curva de eficacia diagnóstica y la clasificación con una
matriz de confusión.
Deep Learning on TB CXR photographs i
R É S U M É
d’efficacité du receveur (ROC) et la classification par
matrice de confusion.
R É S U L T A T S : La cohorte a inclus 138 patients
coinfectés par la TB et le virus de l’immunodéficience
humaine, d’âge médian 34 ans (intervalle interquartile
28–40), dont 54 ont été des femmes. La localisation des
zones pathologiques a été excellente (zone sous la courbe
ROC 0,82). Le logiciel a été capable de distinguer
parfaitement les épanchements pleuraux des
modifications intra-parenchymateuses. Les erreurs de
classification les plus fréquentes ont été les
consolidations prises pour des cavernes et les miliaires
confondues avec un profil interstitiel (et vice versa).
C O N C L U S I O N : L’apprentissage de l’analyse de
photographies de CXR est un outil prometteur.
Davantage d’efforts sont requis pour créer des
ensembles de données plus vastes et de bonne qualité
et aboutir à une meilleure performance de diagnostic.
RESUMEN
R E S U L T A D O S: La cohorte incluy ó 138 pacientes
aquejados de coinfecci ón por el virus de la
inmunodeficiencia humana y TB, con una edad
mediana de 34 años (amplitud intercuartil 28–40).
Cincuenta y cuatro pacientes eran de sexo femenino.
La calidad de la localización de las zonas patológicas fue
excelente (área bajo la curva de eficacia diagnóstica
0,82). El programa informático pudo diferenciar
exactamente el derrame pleural de los cambios
intraparenquimatosos. Los errores más frecuentes
fueron clasificar las cavernas como consolidaciones y
el patrón miliar como un patrón intersticial (y viceversa).
C O N C L U S I Ó N: El análisis por aprendizaje profundo de
las fotos de CXR aparece como un instrumento
prometedor. Se precisan nuevas iniciativas que
contribuyan a crear conjuntos de datos más extensos
de gran calidad y lograr un mejor rendimiento
diagnóstico.