Content uploaded by Katie Shpanskaya
Author content
All content in this area was uploaded by Katie Shpanskaya on Dec 14, 2017
Content may be subject to copyright.
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays
with Deep Learning
Pranav Rajpurkar * 1 Jeremy Irvin * 1 Kaylie Zhu 1Brandon Yang 1Hershel Mehta 1
Tony Duan 1Daisy Ding 1Aarti Bagul 1Curtis Langlotz 2Katie Shpanskaya 2
Matthew P. Lungren 2Andrew Y. Ng 1
Abstract
We develop an algorithm that can detect
pneumonia from chest X-rays at a level ex-
ceeding practicing radiologists. Our algo-
rithm, CheXNet, is a 121-layer convolutional
neural network trained on ChestX-ray14, cur-
rently the largest publicly available chest X-
ray dataset, containing over 100,000 frontal-
view X-ray images with 14 diseases. Four
practicing academic radiologists annotate a
test set, on which we compare the perfor-
mance of CheXNet to that of radiologists.
We find that CheXNet exceeds average radi-
ologist performance on pneumonia detection
on both sensitivity and specificity. We extend
CheXNet to detect all 14 diseases in ChestX-
ray14 and achieve state of the art results on
all 14 diseases.
1. Introduction
More than 1 million adults are hospitalized with pneu-
monia and around 50,000 die from the disease every
year in the US alone (CDC,2017). Chest X-rays
are currently the best available method for diagnosing
pneumonia (WHO,2001), playing a crucial role in clin-
ical care (Franquet,2001) and epidemiological studies
(Cherian et al.,2005). However, detecting pneumo-
nia in chest X-rays is a challenging task that relies on
the availability of expert radiologists. In this work, we
present a model that can automatically detect pneu-
monia from chest X-rays at a level exceeding practicing
radiologists.
*Equal contribution 1Stanford University De-
partment of Computer Science 2Stanford University
School of Medicine. Correspondence to: Pranav Ra-
jpurkar <pranavsr@cs.stanford.edu>, Jeremy Irvin
<jirvin16@cs.stanford.edu>.
Project website at https://stanfordmlgroup.
github.io/projects/chexnet
Output
Pneumonia Positive (85%)
Input
Chest X-Ray Image
CheXNet
121-layer CNN
Figure 1. CheXNet is a 121-layer convolutional neural net-
work that takes a chest X-ray image as input, and outputs
the probability of a pathology. On this example, CheXnet
correctly detects pneumonia and also localizes areas in the
image most indicative of the pathology.
Our model, ChexNet (shown in Figure 1), is a 121-
layer convolutional neural network that inputs a chest
X-ray image and outputs the probability of pneumonia
along with a heatmap localizing the areas of the im-
age most indicative of pneumonia. We train CheXNet
on the recently released ChestX-ray14 dataset (Wang
et al.,2017), which contains 112,120 frontal-view chest
X-ray images individually labeled with up to 14 differ-
ent thoracic diseases, including pneumonia. We use
arXiv:1711.05225v2 [cs.CV] 25 Nov 2017
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
Figure 2. CheXNet outperforms the average of the radiologists at pneuomonia detection using X-ray images. CheXNet
is tested against 4 practicing radiologists on sensitivity (which measures the proportion of positives that are correctly
identified as such) and specificity (which measures the proportion of negatives that are correctly identified as such). A
single radiologist’s performance is represented by an orange marker, while the average is represented by green. CheXNet
outputs the probability of detecting pneumonia in a Chest X-ray, and the blue curve is generated by varying the thresholds
used for the classification boundary. The sensitivity-specificity point for each radiologist and for the average lie below the
blue curve, signifying that CheXNet is able to detect pneumonia at a level matching or exceeding radiologists.
dense connections (Huang et al.,2016) and batch nor-
malization (Ioffe & Szegedy,2015) to make the opti-
mization of such a deep network tractable.
Detecting pneumonia in chest radiography can be diffi-
cult for radiologists. The appearance of pneumonia in
X-ray images is often vague, can overlap with other di-
agnoses, and can mimic many other benign abnormal-
ities. These discrepancies cause considerable variabil-
ity among radiologists in the diagnosis of pneumonia
(Neuman et al.,2012;Davies et al.,1996;Hopstaken
et al.,2004). To estimate radiologist performance, we
collect annotations from four practicing academic radi-
ologists on a subset of 420 images from ChestX-ray14.
On these 420 images, we measure performance of in-
dividual radiologists using the majority vote of other
radiologists as ground truth, and similarly measure
model performance.
We find that the model exceeds the average radiolo-
gist performance at the pneumonia detection task on
both sensitivity and specificity. To compare CheXNet
against previous work using ChestX-ray14, we make
simple modifications to CheXNet to detect all 14 dis-
eases in ChestX-ray14, and find that we outperform
best published results on all 14 diseases. Automated
detection of diseases from chest X-rays at the level of
expert radiologists would not only have tremendous
benefit in clinical settings, it would also be invaluable
in delivery of health care to populations with inade-
quate access to diagnostic imaging specialists.
2. CheXNet
2.1. Problem Formulation
The pneumonia detection task is a binary classification
problem, where the input is a frontal-view chest X-
ray image Xand the output is a binary label y∈
{0,1}indicating the absence or presence of pneumonia
respectively. For a single example in the training set,
we optimize the weighted binary cross entropy loss
L(X, y) = −w+·ylog p(Y= 1|X)
−w−·(1 −y) log p(Y= 0|X),
where p(Y=i|X) is the probability that the network
assigns to the label i,w+=|N|/(|P|+|N|), and w−=
|P|/(|P|+|N|) with |P|and |N|the number of positive
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
cases and negative cases of pneumonia in the training
set respectively.
2.2. Model Architecture and Training
CheXNet is a 121-layer Dense Convolutional Net-
work (DenseNet) (Huang et al.,2016) trained on the
ChestX-ray 14 dataset. DenseNets improve flow of in-
formation and gradients through the network, making
the optimization of very deep networks tractable. We
replace the final fully connected layer with one that
has a single output, after which we apply a sigmoid
nonlinearity.
The weights of the network are initialized with weights
from a model pretrained on ImageNet (Deng et al.,
2009). The network is trained end-to-end using Adam
with standard parameters (β1= 0.9 and β2= 0.999)
(Kingma & Ba,2014). We train the model using mini-
batches of size 16. We use an initial learning rate of
0.001 that is decayed by a factor of 10 each time the
validation loss plateaus after an epoch, and pick the
model with the lowest validation loss.
3. Data
3.1. Training
We use the ChestX-ray14 dataset released by Wang
et al. (2017) which contains 112,120 frontal-view X-ray
images of 30,805 unique patients. Wang et al. (2017)
annotate each image with up to 14 different thoracic
pathology labels using automatic extraction methods
on radiology reports. We label images that have pneu-
monia as one of the annotated pathologies as positive
examples and label all other images as negative exam-
ples. For the pneumonia detection task, we randomly
split the dataset into training (28744 patients, 98637
images), validation (1672 patients, 6351 images), and
test (389 patients, 420 images). There is no patient
overlap between the sets.
Before inputting the images into the network, we
downscale the images to 224×224 and normalize based
on the mean and standard deviation of images in the
ImageNet training set. We also augment the training
data with random horizontal flipping.
3.2. Test
We collected a test set of 420 frontal chest X-rays. An-
notations were obtained independently from four prac-
ticing radiologists at Stanford University, who were
asked to label all 14 pathologies in Wang et al. (2017).
The radiologists had 4, 7, 25, and 28 years of experi-
ence, and one of the radiologists is a sub-specialty fel-
lowship trained thoracic radiologist. Radiologists did
not have access to any patient information or knowl-
edge of disease prevalence in the data. Labels were
entered into a standardized data entry program.
4. CheXNet vs. Radiologist
Performance
4.1. Comparison
We assess radiologist performance on the test set on
the pneumonia detection task. Recall that each of the
images in the test set has a ground truth label from
four practicing radiologists. We evaluate the perfor-
mance of an individual radiologist by using the ma-
jority vote of the other three radiologists as ground
truth. Similarly, we evaluate CheXNet using the ma-
jority vote of three of four radiologists, repeated four
times to cover all groups of three.
We compare CheXNet against radiologists on the Re-
ceiver Operating Characteristic (ROC) curve, which
plots model sensitivity against 1 - specificity. Figure 2
illustrates the model ROC curve as well as the four
radiologist and average radiologist operating points: a
single radiologist’s performance is represented by an
orange marker, while the average is represented by
green. CheXNet outputs the probability of detect-
ing pneumonia in a Chest X-ray, and the ROC curve
is generated by varying the thresholds used for the
classification boundary. CheXNet has an AUROC of
0.828 on the test set. The sensitivity-specificity point
for each radiologist and for the average lie below the
ROC curve, signifying that CheXNet is able to detect
pneumonia at a level matching or exceeding radiolo-
gists.
4.2. Limitations
We identify three limitations of this comparison. First,
neither the model nor the radiologists were permit-
ted to use prior examinations or patient history, which
have been shown to decrease radiologist performance
(Berbaum et al.,1985;Potchen et al.,1979). Second,
only frontal radiographs were presented to the radi-
ologists and model during diagnosis, but it has been
shown that up to 15% of accurate diagnoses require the
lateral view (Raoof et al.,2012); we thus expect that
this setup provides a conservative estimate of perfor-
mance. Third, neither the model nor the radiologists
were not permitted to use patient history, which has
been shown to decrease radiologist diagnostic perfor-
mance in interpreting chest radiographs (for example,
given a pulmonary abnormality with a history of fever
and cough, pneumonia would be appropriate rather
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
Pathology Wang et al. (2017) Yao et al. (2017) CheXNet (ours)
Atelectasis 0.716 0.772 0.8094
Cardiomegaly 0.807 0.904 0.9248
Effusion 0.784 0.859 0.8638
Infiltration 0.609 0.695 0.7345
Mass 0.706 0.792 0.8676
Nodule 0.671 0.717 0.7802
Pneumonia 0.633 0.713 0.7680
Pneumothorax 0.806 0.841 0.8887
Consolidation 0.708 0.788 0.7901
Edema 0.835 0.882 0.8878
Emphysema 0.815 0.829 0.9371
Fibrosis 0.769 0.767 0.8047
Pleural Thickening 0.708 0.765 0.8062
Hernia 0.767 0.914 0.9164
Table 1. CheXNet outperforms the best published results on all 14 pathologies in the ChestX-ray14 dataset. In detecting
Mass, Nodule, Pneumonia, and Emphysema, CheXNet has a margin of >0.05 AUROC over previous state of the art
results.
than less specific terms such as infiltration or consoli-
dation) (Potchen et al.,1979).
5. ChexNet vs. Previous State of the
Art on the ChestX-ray14 Dataset
We extend the algorithm to classify multiple thoracic
pathologies by making three changes. First, instead of
outputting one binary label, ChexNet outputs a vec-
tor tof binary labels indicating the absence or presence
of each of the following 14 pathology classes: Atelec-
tasis, Cardiomegaly, Consolidation, Edema, Effusion,
Emphysema, Fibrosis, Hernia, Infiltration, Mass, Nod-
ule, Pleural Thickening, Pneumonia, and Pneumotho-
rax. Second, we replace the final fully connected layer
in CheXNet with a fully connected layer producing a
14-dimensional output, after which we apply an ele-
mentwise sigmoid nonlinearity. The final output is the
predicted probability of the presence of each pathology
class. Third, we modify the loss function to optimize
the sum of unweighted binary cross entropy losses
L(X, y) =
14
X
c=1
[−yclog p(Yc= 1|X)
−(1 −yc) log p(Yc= 0|X)],
where p(Yc= 1|X) is the predicted probability that
the image contains the pathology cand p(Yc= 0|X)
is the predicted probability that the image does not
contain the pathology c.
We randomly split the dataset into training (70%), val-
idation (10%), and test (20%) sets, following previous
work on ChestX-ray14 (Wang et al.,2017;Yao et al.,
2017). We ensure that there is no patient overlap be-
tween the splits. We compare the per-class AUROC of
the model against the previous state of the art held by
Yao et al. (2017) on 13 classes and Wang et al. (2017)
on the remaining 1 class.
We find that CheXNet achieves state of the art results
on all 14 pathology classes. Table 1illustrates the per-
class AUROC comparison on the test set. On Mass,
Nodule, Pneumonia, and Emphysema, we outperform
previous state of the art considerably (>0.05 increase
in AUROC).
6. Model Interpretation
To interpret the network predictions, we also produce
heatmaps to visualize the areas of the image most in-
dicative of the disease using class activation mappings
(CAMs) (Zhou et al.,2016). To generate the CAMs,
we feed an image into the fully trained network and
extract the feature maps that are output by the final
convolutional layer. Let fkbe the kth feature map and
let wc,k be the weight in the final classification layer
for feature map kleading to pathology c. We obtain
a map Mcof the most salient features used in classi-
fying the image as having pathology cby taking the
weighted sum of the feature maps using their associ-
ated weights. Formally,
Mc=X
k
wc,kfk.
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
(a) Patient with multifocal com-
munity acquired pneumonia. The
model correctly detects the airspace
disease in the left lower and right up-
per lobes to arrive at the pneumonia
diagnosis.
(b) Patient with a left lung nodule.
The model identifies the left lower
lobe lung nodule and correctly clas-
sifies the pathology.
(c) Patient with primary lung ma-
lignancy and two large masses, one
in the left lower lobe and one in
the right upper lobe adjacent to the
mediastinum. The model correctly
identifies both masses in the X-ray.
(d) Patient with a right-sided pneu-
mothroax and chest tube. The
model detects the abnormal lung
to correctly predict the presence of
pneumothorax (collapsed lung).
(e) Patient with a large right pleural
effusion (fluid in the pleural space).
The model correctly labels the effu-
sion and focuses on the right lower
chest.
(f) Patient with congestive heart
failure and cardiomegaly (enlarged
heart). The model correctly identi-
fies the enlarged cardiac silhouette.
Figure 3. CheXNet localizes pathologies it identifies using Class Activation Maps, which highlight the areas of the X-ray
that are most important for making a particular pathology classification. The captions for each image are provided by
one of the practicing radiologists.
We identify the most important features used by the
model in its prediction of the pathology cby upscal-
ing the map Mcto the dimensions of the image and
overlaying the image.
Figure 3shows several examples of CAMs on the pneu-
monia detection task as well as the 14-class pathology
classification task.
7. Related Work
Recent advancements in deep learning and large
datasets have enabled algorithms to surpass the per-
formance of medical professionals in a wide variety of
medical imaging tasks, including diabetic retinopathy
detection (Gulshan et al.,2016), skin cancer classifica-
tion (Esteva et al.,2017), arrhythmia detection (Ra-
jpurkar et al.,2017), and hemorrhage identification
(Grewal et al.,2017).
Automated diagnosis from chest radiographs has re-
ceived increasing attention with algorithms for pul-
monary tuberculosis classification (Lakhani & Sun-
daram,2017) and lung nodule detection (Huang et al.,
2017). Islam et al. (2017) studied the performance
of various convolutional architectures on different ab-
normalities using the publicly available OpenI dataset
(Demner-Fushman et al.,2015). Wang et al. (2017)
released ChestX-ray-14, an order of magnitude larger
than previous datasets of its kind, and also bench-
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
marked different convolutional neural network archi-
tectures pre-trained on ImageNet. Recently Yao et al.
(2017) exploited statistical dependencies between la-
bels in order make more accurate predictions, outper-
forming Wang et al. (2017) on 13 of 14 classes.
8. Conclusion
Pneumonia accounts for a significant proportion of
patient morbidity and mortality (Gon¸calves-Pereira
et al.,2013). Early diagnosis and treatment of pneu-
monia is critical to preventing complications including
death (Aydogdu et al.,2010). With approximately 2
billion procedures per year, chest X-rays are the most
common imaging examination tool used in practice,
critical for screening, diagnosis, and management of a
variety of diseases including pneumonia (Raoof et al.,
2012). However, two thirds of the global population
lacks access to radiology diagnostics, according to an
estimate by the World Health Organization (Mollura
et al.,2010). There is a shortage of experts who can in-
terpret X-rays, even when imaging equipment is avail-
able, leading to increased mortality from treatable dis-
eases (Kesselman et al.,2016).
We develop an algorithm which exceeds the perfor-
mance of radiologists in detecting pneumonia from
frontal-view chest X-ray images. We also show that
a simple extension of our algorithm to detect multi-
ple diseases outperforms previous state of the art on
ChestX-ray14, the largest publicly available chest X-
ray dataset. With automation at the level of experts,
we hope that this technology can improve healthcare
delivery and increase access to medical imaging ex-
pertise in parts of the world where access to skilled
radiologists is limited.
9. Acknowledgements
We would like to acknowledge the Stanford Center for
Artificial Intelligence in Medicine and Imaging for clin-
ical dataset infrastructure support (AIMI.stanford.
edu).
References
Aydogdu, M, Ozyilmaz, E, Aksoy, Handan, Gursel,
G, and Ekim, Numan. Mortality prediction in
community-acquired pneumonia requiring mechan-
ical ventilation; values of pneumonia and intensive
care unit severity scores. Tuberk Toraks, 58(1):25–
34, 2010.
Berbaum, K, Franken Jr, EA, and Smith, WL. The
effect of comparison films upon resident interpre-
tation of pediatric chest radiographs. Investigative
radiology, 20(2):124–128, 1985.
CDC, 2017. URL https://www.cdc.gov/features/
pneumonia/index.html.
Cherian, Thomas, Mulholland, E Kim, Carlin, John B,
Ostensen, Harald, Amin, Ruhul, Campo, Mar-
garet de, Greenberg, David, Lagos, Rosanna,
Lucero, Marilla, Madhi, Shabir A, et al. Standard-
ized interpretation of paediatric chest radiographs
for the diagnosis of pneumonia in epidemiological
studies. Bulletin of the World Health Organization,
83(5):353–359, 2005.
Davies, H Dele, Wang, Elaine E-l, Manson, David,
Babyn, Paul, and Shuckett, Bruce. Reliability of
the chest radiograph in the diagnosis of lower res-
piratory infections in young children. The Pediatric
infectious disease journal, 15(7):600–604, 1996.
Demner-Fushman, Dina, Kohli, Marc D, Rosenman,
Marc B, Shooshan, Sonya E, Rodriguez, Laritza,
Antani, Sameer, Thoma, George R, and McDonald,
Clement J. Preparing a collection of radiology ex-
aminations for distribution and retrieval. Journal of
the American Medical Informatics Association, 23
(2):304–310, 2015.
Deng, Jia, Dong, Wei, Socher, Richard, Li, Li-Jia, Li,
Kai, and Fei-Fei, Li. Imagenet: A large-scale hier-
archical image database. In Computer Vision and
Pattern Recognition, 2009. CVPR 2009. IEEE Con-
ference on, pp. 248–255. IEEE, 2009.
Esteva, Andre, Kuprel, Brett, Novoa, Roberto A,
Ko, Justin, Swetter, Susan M, Blau, Helen M, and
Thrun, Sebastian. Dermatologist-level classification
of skin cancer with deep neural networks. Nature,
542(7639):115–118, 2017.
Franquet, T. Imaging of pneumonia: trends and algo-
rithms. European Respiratory Journal, 18(1):196–
208, 2001.
Gon¸calves-Pereira, Jo˜ao, Concei¸c˜ao, Catarina, and
P´ovoa, Pedro. Community-acquired pneumo-
nia: identification and evaluation of nonresponders.
Therapeutic advances in infectious disease, 1(1):5–
17, 2013.
Grewal, Monika, Srivastava, Muktabh Mayank, Ku-
mar, Pulkit, and Varadarajan, Srikrishna. Radnet:
Radiologist level accuracy using deep learning for
hemorrhage detection in ct scans. arXiv preprint
arXiv:1710.04934, 2017.
CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
Gulshan, Varun, Peng, Lily, Coram, Marc, Stumpe,
Martin C, Wu, Derek, Narayanaswamy, Arunacha-
lam, Venugopalan, Subhashini, Widner, Kasumi,
Madams, Tom, Cuadros, Jorge, et al. Development
and validation of a deep learning algorithm for de-
tection of diabetic retinopathy in retinal fundus pho-
tographs. Jama, 316(22):2402–2410, 2016.
Hopstaken, RM, Witbraad, T, Van Engelshoven,
JMA, and Dinant, GJ. Inter-observer variation in
the interpretation of chest radiographs for pneumo-
nia in community-acquired lower respiratory tract
infections. Clinical radiology, 59(8):743–752, 2004.
Huang, Gao, Liu, Zhuang, Weinberger, Kilian Q, and
van der Maaten, Laurens. Densely connected convo-
lutional networks. arXiv preprint arXiv:1608.06993,
2016.
Huang, Peng, Park, Seyoun, Yan, Rongkai, Lee,
Junghoon, Chu, Linda C, Lin, Cheng T, Hussien,
Amira, Rathmell, Joshua, Thomas, Brett, Chen,
Chen, et al. Added value of computer-aided ct image
features for early lung cancer diagnosis with small
pulmonary nodules: A matched case-control study.
Radiology, pp. 162725, 2017.
Ioffe, Sergey and Szegedy, Christian. Batch normaliza-
tion: Accelerating deep network training by reduc-
ing internal covariate shift. In International Confer-
ence on Machine Learning, pp. 448–456, 2015.
Islam, Mohammad Tariqul, Aowal, Md Abdul, Min-
haz, Ahmed Tahseen, and Ashraf, Khalid. Ab-
normality detection and localization in chest x-rays
using deep convolutional neural networks. arXiv
preprint arXiv:1705.09850, 2017.
Kesselman, Andrew, Soroosh, Garshasb, Mollura,
Daniel J, and Group, RAD-AID Conference Writ-
ing. 2015 rad-aid conference on international radi-
ology for developing countries: The evolving global
radiology landscape. Journal of the American Col-
lege of Radiology, 13(9):1139–1144, 2016.
Kingma, Diederik and Ba, Jimmy. Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Lakhani, Paras and Sundaram, Baskaran. Deep learn-
ing at chest radiography: Automated classification
of pulmonary tuberculosis by using convolutional
neural networks. Radiology, pp. 162326, 2017.
Mollura, Daniel J, Azene, Ezana M, Starikovsky,
Anna, Thelwell, Aduke, Iosifescu, Sarah, Kimble,
Cary, Polin, Ann, Garra, Brian S, DeStigter, Kris-
ten K, Short, Brad, et al. White paper report of
the rad-aid conference on international radiology for
developing countries: identifying challenges, oppor-
tunities, and strategies for imaging services in the
developing world. Journal of the American College
of Radiology, 7(7):495–500, 2010.
Neuman, Mark I, Lee, Edward Y, Bixby, Sarah,
Diperna, Stephanie, Hellinger, Jeffrey, Markowitz,
Richard, Servaes, Sabah, Monuteaux, Michael C,
and Shah, Samir S. Variability in the interpretation
of chest radiographs for the diagnosis of pneumo-
nia in children. Journal of hospital medicine, 7(4):
294–298, 2012.
Potchen, EJ, Gard, JW, Lazar, P, Lahaie, P, and
Andary, M. Effect of clinical history data on chest
film interpretation-direction or distraction. In Inves-
tigative Radiology, volume 14, pp. 404–404, 1979.
Rajpurkar, Pranav, Hannun, Awni Y, Haghpanahi,
Masoumeh, Bourn, Codie, and Ng, Andrew Y.
Cardiologist-level arrhythmia detection with con-
volutional neural networks. arXiv preprint
arXiv:1707.01836, 2017.
Raoof, Suhail, Feigin, David, Sung, Arthur,
Raoof, Sabiha, Irugulpati, Lavanya, and Rosenow,
Edward C. Interpretation of plain chest
roentgenogram. CHEST Journal, 141(2):545–558,
2012.
Wang, Xiaosong, Peng, Yifan, Lu, Le, Lu, Zhiyong,
Bagheri, Mohammadhadi, and Summers, Ronald M.
Chestx-ray8: Hospital-scale chest x-ray database
and benchmarks on weakly-supervised classification
and localization of common thorax diseases. arXiv
preprint arXiv:1705.02315, 2017.
WHO. Standardization of interpretation of chest ra-
diographs for the diagnosis of pneumonia in chil-
dren. 2001.
Yao, Li, Poblenz, Eric, Dagunts, Dmitry, Covington,
Ben, Bernard, Devon, and Lyman, Kevin. Learning
to diagnose from scratch by exploiting dependen-
cies among labels. arXiv preprint arXiv:1710.10501,
2017.
Zhou, Bolei, Khosla, Aditya, Lapedriza, Agata, Oliva,
Aude, and Torralba, Antonio. Learning deep fea-
tures for discriminative localization. In Proceedings
of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 2921–2929, 2016.
