=Paper=
{{Paper
|id=Vol-2696/paper_99
|storemode=property
|title=UAIC2020: Lung Analysis for Tuberculosis Detection
|pdfUrl=https://ceur-ws.org/Vol-2696/paper_99.pdf
|volume=Vol-2696
|authors=Lucia Georgiana Coca,Alexandra Hanganu,Ciprian-Gabriel Cusmuliuc,Adrian Iftene
|dblpUrl=https://dblp.org/rec/conf/clef/CocaHCI20
}}
==UAIC2020: Lung Analysis for Tuberculosis Detection==
https://ceur-ws.org/Vol-2696/paper_99.pdf
UAIC2020: Lung Analysis for Tuberculosis Detection
Lucia-Georgiana Coca, Alexandra Hanganu, Ciprian-Gabriel Cusmuliuc, Adrian Iftene
“Alexandru Ioan Cuza” University, Faculty of Computer Science, Iasi, Romania
{coca.lucia.georgiana, alexandra.hanganu,
cusmuliuc.ciprian.gabriel, adiftene}@info.uaic.ro
Abstract. Tuberculosis is a bacterial infection that is transmitted through the air
and affects the lungs. According to a WHO report, in Europe, in 2016 there
were almost 228,000 people infected with tuberculosis. This disease can be
treated if diagnosed in time and followed by a CT scan. For the 4th edition of
the ImageCLEFmed Tuberculosis task, organizers proposed a new approach
with a major impact on the real-world clinical routines, regarding generation
lung-based rather than CT-based. Our solution proposed for generating an au-
tomated CT report is based on Support Vector Machine and Convolutional Neu-
ral Networks. The first algorithm we used is SVM, which provided the best re-
sults for the test images. Furthermore, CNN was also used and provided results
that were almost as accurate as SVM.
Keywords: Support Vector Machine, Tuberculosis, CNN.
1 Introduction
Tuberculosis (TB) is caused by bacteria (Mycobacterium tuberculosis) that most often
affects the lungs. Tuberculosis is curable and preventable if it is discovered and diag-
nosed in time. TB is spread from person to person through the air. When people with
lung TB cough, sneeze or spit, they propel the TB germs into the air. A person needs
to inhale only a few of these germs to become infected, due to the large number of
people infected with TB, many researchers are involved in finding better solutions in
treating and maintaining TB patients.
ImageCLEF 2020 [1] is an evaluation campaign that is being organized as part of
the CLEF initiative labs1. In this year's edition, the organizers decided to focus on the
task of automatically generating a CT report, as it has important results that can have
a major impact on real-world clinical routines. To make the task as attractive as pos-
sible, this year the generation of reports is based on the lungs and not on CTs. Labels
for the left and right lungs were provided independently. The set of target labels in the
CT Report have been updated in accordance with the opinion of the medical experts.
This year, 3 labels were offered for each lung: the presence of tuberculosis lesions in
1 http://www.clef-initiative.eu/
Copyright © 2020 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0). CLEF 2020, 22-25 Septem-
ber 2020, Thessaloniki, Greece.
general, the presence of pleurisy and especially of the caverns. The size of the data set
was also increased compared to the previous year.
This paper describes the participation of team UAIC2020, from the Faculty of
Computer Science, “Alexandru Ioan Cuza” University of Iasi, in ImageCLEFtubercu-
losis 2020 task [2], at ImageCLEF 2020. The remaining of this paper was organized
as follows: Section 2 describes the state of the art, Section 3 presents our methods,
Section 4 evaluates the methods proposed and, in the end, we draw a conclusion and
present future work.
2 State of the art
In previous editions, multiple different approaches were proposed for this task, mainly
based on SVMs and CNNs.
In 2018 the most notable result was of team UIIP_BioMed [3] that managed to ob-
tain the highest kappa score, 0.2312, with an accuracy of 0.4227 and root mean
squared error of 0.7840. Furthermore, the team MedGift [4] scored 0.7708 in terms of
ROC AUC, the highest in that year. Both teams had very different approaches,
UIIP_BioMed used a Deep CNN and MedGift had an SVM with and RBF kernel.
In 2019 the leaderboard has not changed much, UIIP_BioMed [5] is still the leader
with a mean AUC of 0.7968 and a min AUC of 0.6860 followed by CompElecEngCU
[6] with an AUC of 0.7629. UIIP_BioMed used a 2D CNN whilst CompElecEngCU
created 2D derived images by concatenating sagittal and coronal CT slices that are
classified with a hybrid of a 2D CNN based on AlexNet [7]. Another approach, FI-
IAugt [8], uses group performed random sampling of pixels of the CT volumes and
used a combination of decision trees and weak classifiers in order to participate in the
SVR task, this team is from our University and ranked 38 out of 54.
Our experience with classification and predictions stems from other similar works
such as predicting cracks in images using CNN [9] but also fake news identification
[10], anorexia detection [11] and claim identification [12].
3 Methods
The solution proposed for generating an automated CT report is based on Support
Vector Machine and Convolutional Neural Networks.
Regarding the data, we converted the 3D NIfTI2 file format to 2D PNGs represent-
ing specific sides of the lung. This transformation also had to be done to the masks
provided by the organizers [13] [14] as they needed to match the previously resulted
CT scans. In terms of pre-processing, the mask images were divided by lung lobes
and the background was changed to black whilst the lung color was set to white. By
doing so, new masks were generated for each lung, enabling us to work with each of
them.
2 https://radiopaedia.org/articles/nifti-file-format
The first algorithm we used is SVM [15], which provided the best results for the
test images. Furthermore, CNN [16] was also used and provided results that were
almost as accurate as SVM. The work was done in parallel and all the findings from
one test run were tested on the other, leading both to a final form. The two algorithms
dealt with each area and lung separately (each lung had a correspondent in the code
for the Affected column, the Caverns column and the Pleurisy column), leading to a
more complex solution that was catered to fit the needs of the given task. The final
CT report was computed individually for both algorithms and the results were com-
pared after each hyperparameter tuning, ensuring the best result.
3.1 Dataset
The given NIfTI files for the train images also had a CT Report, which was extremely
helpful when training an algorithm as one could label the data accordingly. The da-
taset was highly imbalanced and that could have had a negative impact on the training
of an algorithm.
Table 1 provides an insight in the patient affected in the training set.
Table 1. Patients Affected in the Training Set
Columns Patients Affected in the Training Set (out of 283)
LeftLungAfftected 211
RightLungAffected 233
CavernsLeft 66
CavernsRight 79
PleurisyLeft 7
PleurisyRight 14
3.2 Architecture
This section contains a description of the methods used for obtaining the results. The
overall architecture of the project used to generate the CT Report is depicted from
Figure 1. It can be seen that the generated CT is fed to each algorithm individually.
The algorithms use the masks applied on the CT scans in order to issue a response.
3.3 Data Preprocessing
The solutions proposed required the images in NIfTI file format to be transformed
into a 2D format. To achieve this, the file has been loaded using the NiBabel library3
and then transformed into a NumPy4 array. This array was then transformed to respect
3 https://nipy.org/nibabel/
4 https://numpy.org/
the type numpy.int165, having the shape equal to 3. This meant that there would be
three available ways for the slices to be made: from left to right, from front to back
and from top to bottom. Initially, 512 slices were made for each region.
Fig. 1. Architecture of the model.
Working with the initial results of such a slicing led to a different approach for each
of the regions. For the slices made from left to right, only a quarter of them remained,
as they tended to become repetitive and the data was ruled redundant; moreover, the
first 5 and last 10 pictures were removed as they oftentimes contained little to no in-
formation for our algorithms. The entirely empty pictures were eliminated based on
the fact that the matrix was filled with the value 255 and offered no relevant infor-
mation for further use. However, for slices from the front of the lungs to the back, the
delete process was done in batches. First one was from 0 to 241, where the black im-
ages were removed and only a quarter of the remaining slices were used for working
further with them. The second batch was from 241 to 272. Here, all slices were kept
as these have proven to be the ones with the best resolution, highest contrast and most
information for our algorithms. The last batch, was the one which contained images
from 272 to 512, which is the total number of slices and its’ processing consisted in
the removal of the black images, while keeping only a quarter of the number of slices
for further work. Finally, for the top to bottom resulting slices, only a quarter re-
mained, the first 4 and last 8 pictures were removed and the entirely empty pictures
were eliminated also. After selecting the required slices from both the CT scans and
masks, a decision was taken to separate the lungs. Using the first set of masks, as the
color-coded lungs were easier to work with, the images have been transformed into
NumPy arrays and the work was done on the resulting array. For the anatomically left
lung, the 127 values were set to 0 and for the right lung, 255 values were transformed
into 0 and 127 values to 255. The masks that have resulted from this preprocessing
5 https://numpy.org/doc/stable/reference/generated/numpy.dtype.html
were then separated so work could be done on each lung lobe, individually for the
three types of affections.
Fig. 2. Image preprocessing example.
Table 2. Dataset before and after preprocessing.
Images Train Test
Before After Before After
CT Scan Images 455117 47892 184320 21951
Mask Images 477117 73762 184320 31552
3.4 Mask Application
After obtaining the images for both CT Scans and Masks, the next step was to apply
the masks onto the scans6. To do this, the cv27 library was used.
The method was based on taking a mask image and finding its correspondent in the
CT Scan images. Our approach was to apply the images on the masks and not the
other way around as to ensure that there would be no errors. After using cv2.imread 8
to load the images, the masks had to be resized and for that we used as arguments the
loaded mask image and for shape the scan image’s shape[1::-1]. The last step was to
use the bitwise_and9 function to apply the mask onto the image and then
cv2.imwrite10 to save the resulting file.
6 https://note.nkmk.me/en/python-opencv-numpy-alpha-blend-mask/
7 https://opencv.org/
8 https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_gui/py_image_display
/py_image_display.html?highlight=imread
9 https://docs.opencv.org/2.4/modules/core/doc/operations_on_arrays.html#bitwise-and
10 https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_gui/py_image_display
/py_image_display.html?highlight=imwrite
3.5 Implementation
Our team proposed two different types of algorithms for the generation of an automat-
ic CT Report: Support Vector Machine and Convolutional Neural Network. The two
algorithms are different at core and required data to be looked at individually from
this point on.
Fig. 3. Before and after mask application
To fully understand the implications of such an impactful disease and how to imple-
ment these algorithms, several professionals and students working in the medical field
have been asked for insight. They have provided important information about how an
expert looks at a CT scan of a patient with tuberculosis. Moreover, pointers have been
given for important details to keep in mind when examining such an image. This ini-
tial research into the medical field has proved to be of great help. Being told that look-
ing exclusively at the inferior part of a lung to find signs of pleurisy or for missing
areas inside lungs to find caverns, has been a considerable guidance. Hence, this part
of the research has been insightful and of tremendous value when going forward with
the implementation of the two machine learning algorithms. The troubleshooting and
the interpretation of the results has been easier as the problem was understood from
both a technical and medical point of view.
Since SVM is a classification problem at core, the idea behind this implementation
is that, taking all individual pictures from one patient, they were labeled accordingly,
with 1 or 0, and then the final result represented the mean of all values returned for
that patient, for that column. For the training of this algorithm, there were multiple
libraries used, such as NumPy, Joblib11, Keras12 and sklearn13. The data needed to be
transformed into a NumPy array first. For that, it was saved into a three dimensional
11 https://joblib.readthedocs.io/en/latest/
12 https://keras.io/
13 https://scikit-learn.org/stable/
array, that held the side (left or right), the region of the lung (front to back, left to
right or top to bottom) and the type of the illness (affected, caverns, pleurisy) and only
after it was saved into a np.array14. What is to be noted is that the pictures have been
resized to 32×32 pixels, as they were initially 512×512. When it comes to the actual
training, svm.SVC15 was used with the ‘rbf’ kernel and other different parameters that
will be discussed during evaluation. The resulting output was saved using joblib, so
that when conducting further tests, it will be easier to load the data. Computing was
done also onto a three-dimensional array, as it was the best way to hold information
individually. Our training function had an input parameter so that the state of the
models would be available to the algorithm (saved was True if there existed saved
models and False otherwise). The next step was to compute the final values for each
patient and then use them to write the final CT Report.
The CNN algorithm had a different approach to the issue. Each of the columns was
inspected individually, one script for each one, as there was a need to control the data
better and to adapt the parameters for the best possible results. In each one, the ap-
proach is similar. The training dataset provided by the organizers was split into train-
ing and validation subsets for computing the final trained model. The images have
been resized, different values for different outputs and reshaped accordingly. For this
to happen, many libraries needed to be used for the backend: OpenCV (cv2), Numpy,
Pillow16, Keras and Tensorflow17. Our approach was based on individual picture la-
beling and achieving a result for the whole set of pictures for a test patient, by compu-
ting, once again, a mean of the resulted values that, this time were not 0 or 1, but ra-
ther values belonging to this interval. Before the actual training, the images had to be
assigned a label. This was done making use of the CT Report given by the organizers,
and thus each picture from a patient received the corresponding label. Then, the imag-
es were resized and transformed into NumPy arrays, by going through all the pictures
from all of the regions from the same patient. After saving the training data so it can
be used further for testing, the data was shuffled. The test data had a similar process,
only there was no shuffling involved at the end.
For training the model, there were 500 pictures set aside for the validation set.
Then, the arrays were reshaped to accommodate our image size. Moving further,
Conv2D18 was used for adding layers to the model. The number of said layers was not
constant throughout all of the submissions, but they had the activation set to ‘ReLu’
and the kernel size to 3. After, a flatten layer was added and after this one, a new fully
connected layer was the one to end this cycle, the activation used for it being ‘soft-
max’. Following this, the model was compiled using the ‘Adam optimizer [17], the
loss set to ‘categorical_crossentropy' and the metrics equal to ‘accuracy’.
The next and final step from this set was the fitting of the model and, for this, there
were several parameters used, as they vary from one submission to another, but the
batch size is consistently equal to 200. The last step in this solution was writing the
14 https://numpy.org/doc/stable/reference/generated/numpy.array.html
15 https://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html
16 https://pillow.readthedocs.io/en/stable/
17 https://www.tensorflow.org/
18 https://www.tensorflow.org/api_docs/python/tf/keras/layers/Conv2D
findings in individual comma separated values files and then combining all 6 resulted
files to obtain the final result.
4 Evaluation
In this section we will discuss the evaluation of the model and the differences be-
tween each submission.
In official results we ranked 7 out of 9, with submission ID 68081, and total of 67
teams. Table 3 presents each submission results, we made a total of 8 submissions.
Table 3. Official evaluation results
ID mean_aucc min_aucc
67548 0.617 0.537
67549 0.591 0.503
67573 0.609 0.598
67977 0.505 0.402
68081 0.659 0.562
68121 0.626 0.513
68134 0.608 0.513
68135 0.491 0.382
Submissions information:
• ID 67548 – This submission was computed using SVM algorithm, where the
gamma parameter was equal to 0.005 and the C parameter was equal to
1000.0.
• ID 67549 – This submission was computed using SVM algorithm, where
there was no gamma parameter preset and the C parameter is equal to 1.0.
• ID 67573 – This submission was computed using CNN algorithm, where the
image size was set to (64, 64).
• ID 67977 – This submission was computed using CNN algorithm, where the
image size was set to (128, 128).
• ID 68081 – This submission is based on the CNN approach, we tried to im-
prove that algorithm. We used Pre-Activation ResNet with Identity Mapping
[18] with standard parameters; unfortunately, we got marginal result.
• ID 68121 – This submission is based on submission with ID 67548 in an ef-
fort to improve that result. We tried using a different algorithm than SVC, we
used LinearSVC and the parameters were the following: C=15, maximum it-
erations were 1,000 and dual was False.
• ID 68134 – This submission is based on submission with ID 68121. We no-
ticed that the pleurisy for the left lung had very low accuracy so we tried us-
ing a special SVM for that one so we used SVC with ‘rbf’ kernel, C 1000,
gamma 0.005 and cache size 2048; we left the other settings the same as in
ID 68121. We did not get the expected result.
• ID 68135 – This submission was computed using CNN algorithm, where
there were 3 layers in the network (one more than the others before it) and the
image size was set to (64, 64). In this submission the code was supposed to
be adapted to work best on each column of the dataset, so there would be
some of the columns with just 2 layers, or some with only 2 epochs.
We implemented an accuracy calculation mechanism using sklearn.metrics19
roc_curve20, we calculated the accuracy for each side of the lung then we averaged it,
this allowed us to predict the accuracy of our algorithms and analyze the errors with-
out submitting a run.
4.1 Error analysis
The SVM algorithm was the first one to prove that the difference between affected
and unaffected left lobes in patients was a real issue we would have to deal with. This
has made it easier to understand where the issue is and to try and find solutions. The
hyperparameter tuning in the SVM algorithm was done using grid search [19] on 5
values for gamma and C, as follows: ‘gamma’: [1e-3, 1e-4, 10, 100, 1e-5], ‘C’: [1, 10,
100, 1,000, 10,000, 100,000]. Unfortunately, these 30 tests were cut short by the lack
of proper equipment and only provided the first submission parameters. Considering
in [20] SVM had an average accuracy of 0.78 by using a combination of “Gaussian
kernel function with a width parameter of 2.0, while the value of C was set to 10” it is
safe to say there is still room for improvement in future work.
When training the CNN algorithm, oftentimes there was an overfitting problem for
some of the columns due to unbalanced data, also a problem was the fact that long
training times resulted in a lack of hyperparameter tuning possibility. Data imbalance
could not be solved by removing unnecessary data as the result suffered when this
was done. Error reduction can be achieved by creating a more complex architecture as
described in [5] which obtained an accuracy of 0.7968 proving the superiority of this
algorithm over SVM and the potential of neural networks.
5 Conclusion
In this paper we present different ways of detecting tuberculosis and thanks to compe-
titions such as ImageCLEFmed Tuberculosis every year we are a step closer to per-
fecting these techniques.
Our methods are mainly focused around SVM and Neural Networks. We got a
good result that can be improved; thus, we plan on further building on the ResNet
19 https://scikit-learn.org/stable/modules/classes.html#sklearn-metrics-metrics
20https://scikit-learn.org/stable/modules/generated/sklearn.metrics.roc_curve.html#sklearn.
metrics.roc_curve
architecture, improving the current SVM approach and improving our current hard-
ware in order to process the data on GPU for a more streamlines hyperparameter tun-
ing.
Acknowledgements
Special thanks go to our colleagues from second year group E1 without whom this
work would probably have not been possible.
References
1. Bogdan Ionescu, Henning Müller, Renaud Péteri, Asma Ben Abacha, Vivek Datla, Sadid
A. Hasan, Dina Demner-Fushman, Serge Kozlovski, Vitali Liauchuk, Yashin Dicente Cid,
Vassili Kovalev, Obioma Pelka, Christoph M. Friedrich, Alba García Seco de Herrera,
Van-Tu Ninh, Tu-Khiem Le, Liting Zhou, Luca Piras, Michael Riegler, På l. Halvorsen,
Minh-Triet Tran, Mathias Lux, Cathal Gurrin, Duc-Tien Dang-Nguyen, Jon Chamberlain,
Adrian Clark, Antonio Campello, Dimitri Fichou, Raul Berari, Paul Brie, Mihai Dogariu,
Liviu Daniel Ștefan, Mihai Gabriel Constantin, Overview of the ImageCLEF 2020: Mul-
timedia Retrieval in Medical, Lifelogging, Nature, and Internet Applications In: Experi-
mental IR Meets Multilinguality, Multimodality, and Interaction. Proceedings of the 11th
International Conference of the CLEF Association (CLEF 2020), Thessaloniki, Greece,
LNCS Lecture Notes in Computer Science, Springer (September 22-25 2020).
2. Serge Kozlovski, Vitali Liauchuk, Yashin Dicente Cid, Aleh Tarasau, Vassili Kovalev,
Henning Müller, Overview of ImageCLEFtuberculosis 2020 - Automatic CT-based Report
Generation and Tuberculosis Severity Assessment, CLEF working notes, CEUR, 2020.
3. Liauchuk, V., Tarasau, A., Snezhko, E., Kovalev, V. (2018) Imageclef 2018: Lesion-based
tb-descriptor for ct image analysis. In CLEF 2018 Working Notes.
4. Dicente Cid, Y., Müller, H. (2018) Texture-based graph model of the lungs for drug re-
sistance detection, tuberculosis type classification, and severity scoring: Participation in
the imageclef 2018 tuberculosis task. In CLEF 2018 Working Notes.
5. Liauchuk, V. (2019) ImageCLEF 2019: Projection-based CT Image Analysis for TB
Severity Scoring and CT Report Generation. In CLEF2019 Working Notes. 2380, Lugano,
Switzerland, CEUR-WS.org http://ceur-ws.org/Vol-2380 (September 9-12 2019)
6. Mossa, A.A., Yibre, A.M., C¸ Evik, U. (2019) Multi-View CNN with MLP for Diagnosing
Tuberculosis Patients Using CT Scans and Clinically Relevant Metadata. In CLEF2019
Working Notes. 2380 of CEUR Workshop Proceedings., Lugano, Switzerland, CEUR-
WS.org
7. Krizhevsky, A., Sutskever, I., Hinton, G. E. (2017) ImageNet classification with deep con-
volutional neural networks. In Communications of the ACM. 60 (6): 84-90.
doi:10.1145/3065386. ISSN 0001-0782. https://papers.nips.cc/paper/4824-imagenet-
classification-with-deep-convolutional-neural-networks.pdf
8. Tabarcea, A., Rosca, V., Iftene, A. (2019) ImageCLEFmed Tuberculosis 2019: Predicting
CT Scans Severity Scores using Stage-Wise Boosting in Low-Resource Environments. In:
CLEF2019 Working Notes. 2380, Lugano, Switzerland, CEUR-WS.org
9. Coca, G. L., Romanescu, S. C., Botez, S. M., Iftene, A. (2020) Crack detection system in
AWS Cloud using Convolutional neural networks. In 24th International Conference on
Knowledge-Based and Intelligent Information & Engineering Systems (KES 2020).
10. Cușmuliuc, C. G., Coca, L. G., Iftene, A. (2018) Identifying Fake News on Twitter using
Naive Bayes, SVM and Random Forest Distributed Algorithms. In Proceedings of The
13th Edition of the International Conference on Linguistic Resources and Tools for Pro-
cessing Romanian Language (ConsILR-2018). ISSN: 1843-911X, 177-188
11. Cușmuliuc, C. G., Coca, L. G., Iftene, A. (2019) Early Detection of Signs of Anorexia in
Social Media. In 5th Proceedings of the Conference on Mathematical Foundations of In-
formatics. 3-6 July 2019, Iasi, Romania, 245-260.
12. Coca, L. G., Cușmuliuc, C. G., Iftene, A. (2019) CheckThat! 2019 UAICS. In Working
Notes of CLEF 2019 - Conference and Labs of the Evaluation Forum, Lugano, Switzer-
land, September 9-12
13. Yashin Dicente Cid, Oscar A. Jiménez-del-Toro, Adrien Depeursinge, and Henning Mül-
ler. Efficient and fully automatic segmentation of the lungs in CT volumes. In: Goksel, O.,
et al. (eds.) Proceedings of the VISCERAL Challenge at ISBI. No. 1390 in CEUR Work-
shop Proceedings (Apr 2015)
14. Liauchuk, V., Kovalev, V. ImageCLEF 2017: Supervoxels and co-occurrence for tubercu-
losis CT image classification. In: CLEF2017 Working Notes. CEUR Workshop Proceed-
ings, Dublin, Ireland, CEUR-WS.org http://ceur-ws.org (September 11-14 2017) (URL:
http://ceur-ws.org/Vol-1866/paper_146.pdf)
15. Cortes, C., Vapnik, V. N. (1995) Support-vector networks. In Machine Learning. 20 (3):
273–297. CiteSeerX 10.1.1.15.9362. doi:10.1007/BF00994018
16. Valueva, M. V., Nagornov, N. N., Lyakhov, P. A., Valuev, G. V., Chervyakov, N. I.
(2020) Application of the residue number system to reduce hardware costs of the convolu-
tional neural network implementation. In Mathematics and Computers in Simulation.
Elsevier BV. 177: 232–243. doi:10.1016/j.matcom.2020.04.031. ISSN 0378-4754. Convo-
lutional neural networks are a promising tool for solving the problem of pattern recogni-
tion.
17. Kingma, D. P. & Ba, J. (2014), 'Adam: A Method for Stochastic Optimization' , cite
arxiv:1412.6980 Comment: Published as a conference paper at the 3rd International Con-
ference for Learning Representations, San Diego, 2015 .
18. He, Kaiming & Zhang, Xiangyu & Ren, Shaoqing & Sun, Jian. (2016). Identity Mappings
in Deep Residual Networks. 9908. 630-645. 10.1007/978-3-319-46493-0_38.
19. Bergstra, J., Bengio, Y. (2012) Random search for hyper-parameter optimization. In Jour-
nal of Machine Learning Research, 13: 281-305 http://www.jmlr.org/papers/volume13/
bergstra12a/bergstra12a.pdf
20. Nikházy L., Horváth G., Horváth Á., Müller V. (2010) Computer-Aided Detection of
COPD Using Digital Chest Radiographs. In: Bamidis P.D., Pallikarakis N. (eds) XII Medi-
terranean Conference on Medical and Biological Engineering and Computing 2010.
IFMBE Proceedings, vol 29. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-
642-13039-7_63