Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж
The problem of classification of breast tumors on medical images is con-sidered. For its solution the new class of convolutional neural networks-hybrid CNN–FNN network is developed in which convolutional neural network VGG-16 is used as the feature extractor while fuzzy neural network NEFClass is us...
Saved in:
| Date: | 2018 |
|---|---|
| Main Authors: | , , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2018
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/152060 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | System research and information technologies |
| Download file: |  |
Institution
System research and information technologies| _version_ | 1866391910273253376 |
|---|---|
| author | Zaychenko, Yuriy Hamidov, G. Varga, I. |
| author_facet | Zaychenko, Yuriy Hamidov, G. Varga, I. |
| author_sort | Zaychenko, Yuriy |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2019-04-26T15:57:21Z |
| description | The problem of classification of breast tumors on medical images is con-sidered. For its solution the new class of convolutional neural networks-hybrid CNN–FNN network is developed in which convolutional neural network VGG-16 is used as the feature extractor while fuzzy neural network NEFClass is used as the classifier. Training algorithms of FNN were implemented. The experimental investigations of the suggested hybrid network on the standard data set were carried out and comparison with known results was performed. The problem of data dimensionality reduction is considered and application of PCM method is investigated. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2018.4.03 |
| first_indexed | 2025-07-17T10:24:15Z |
| format | Article |
| fulltext |
YU. Zaychenko, G. Hamidov, I. Varga, 2018
Системні дослідження та інформаційні технології, 2018, № 4 37
UDC 683.519
DOI: 10.20535/SRIT.2308-8893.2018.4.03
MEDICAL IMAGES OF BREAST TUMORS DIAGNOSTICS WITH
APPLICATION OF HYBRID CNN–FNN NETWORK
YU. ZAYCHENKO, G. HAMIDOV, I. VARGA
Abstract. The problem of classification of breast tumors on medical images is con-
sidered. For its solution the new class of convolutional neural networks-hybrid
CNN–FNN network is developed in which convolutional neural network VGG-16 is
used as the feature extractor while fuzzy neural network NEFClass is used as the
classifier. Training algorithms of FNN were implemented. The experimental inves-
tigations of the suggested hybrid network on the standard data set were carried out
and comparison with known results was performed. The problem of data dimension-
ality reduction is considered and application of PCM method is investigated.
Keywords: medical diagnostics, breast cancer classification, FNN, CNN, hybrid
network, dimensionality reduction, PCM.
INTRODUCTION. STATE-OF-ART PROBLEM ANALYSIS
Now cancer constitute the great problem for health defense all over the world.
Basing on the data of IARC (International Agency of Cancer Research) 8,2 mil-
lion death cases were registered in year 2012, 27 million new cases of illness are
expected till 2030 [1]. Among the different types of cancer breast cancer takes
the second place by its occurrence in women . Besides, mortality of it is very high
as compared with other cancer diseases [1].
Nowadays, in practice, at every stage of diagnostics information technolo-
gies are utilized. The main goal of medical automated systems are extension of
spheres of practical tasks which may be solved with computers aid, raise of level
intellectual decision support of doctors in particularly in process of express diag-
nostics based on processing and analysis of medical images of human tissue ob-
tained by different source ( MRT, CT etc).
In medical diagnostics problems substantial amount of problem constitute
the features extraction for further processing and the choice of features for classi-
fication method. With development and wide dissemination of decision-support
systems the demands to training algorithms are increasing. Reliability and sim-
plicity of application influence on speed and quality of decision- making which is
very important for express medical diagnostics. The advantages of medical diag-
nostics systems are speed, automation and stability of work which make them
very comfortable tools for express medical diagnostics. Despite of young age of
medical informatics which doesn’t exceed 30 years information technologies in a
whole are fast penetrating in various spheres of medicine and health defence
(family medicine, insurance medicine, building unified information space, inte-
gration in European medical space etc).
Despite of progress which was achieved by diagnostics technologies final
diagnosis of breast cancer including classification of tumours and diagnosis still is
YU. Zaychenko, G. Hamidov, I. Varga
ISSN 1681–6048 System Research & Information Technologies, 2018, № 4 38
performed by pathologist-anatomists which use visual analysis of histological pat-
terns by microscope. The latest achievements in images processing technologies
and machine learning enable to construct systems of automatic detection and di-
agnostics that may help pathologist-anatomists to make true diagnosis and accel-
erate his work. Classification of images histopathology on different patterns
which corresponds to cancer and not-cancer states of tissue is often first rank goal
in images analysis systems for automatic cancer diagnostics.
Up to date several models and methods were developed for breast cancer
detection using various machine learning algorithms. Using such methods and
technologies of AI as neuron networks and SVM [2, 3] accuracy of diagnostics
from 76% to 94% was attained at data set with 92 images.
Zhang and others [4] suggested cascade classifiers approach. At the first cas-
cade level the classifiers reject easy cases (those which evidently don’t pass test)
and the others are transferred to the second level which uses more complex classi-
fication system and so on. This method was applied to data base of Israel techno-
logical Institute consisting of 361 images and accuracy results was 97%. The
most of last papers refers to field of breast cancer classification oriented on inte-
ger images [3–6]. But wide implementation of breast image classification (BIC)
and other forms of digital pathology faces with such disturbances as high cost of
implementation, insufficient productivity for huge amount of clinic procedures,
interior technologic problems, and opposition from pathologist-anatomists side.
Till now the most of works based on histology breast cancer analysis were per-
formed on not large datasets. Some improvement presents data set with 7909
breast images obtained from 82 patients [7]. In this research the authors estimated
various texture descriptors and various classifiers and carried out the experiments
with accuracy from 82% to 85%.
Based on results presented in [7] one can make the conclusion that texture
descriptors may propose good solution for images processing. But some re-
searchers believe that main weakness of modern machine learning methods oc-
curs just at this stage. This means that machine learning algorithms should be less
dependent on functional engineering and be able to extract and organize discrimi-
nating information directly from images, in other words be capable to learn pres-
entations.
The idea of learning presentations isn’t new one but it became implement-
able only now with appearance GPU( Graphic Processing Units) which are capa-
ble to provide high speed performance ( productivity) with relatively small cost
due to their parallel architecture [8].
The alternative to this approach is the application of CNN for medical im-
ages processing and diagnostics, which is considered and developed in the present
research. It was shown that CNN is able to overcome the conventional texture
descriptors [9, 10]. Besides traditional approach to detection of features based on
descriptors demands much efforts and high level knowledge of experts and usu-
ally is specific for every task that prevents its direct application for another simi-
lar tasks.
Therefore in our research we suggested and developed hybrid CNN–FNN
medical images classification system in which CNN is utilized to extract informa-
tive features of images and FNN NEFClass is applied for classification of de-
tected breast tumors on images in two classes: benign and malicious ones.
Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network …
Системні дослідження та інформаційні технології, 2018, № 4 39
The main goal of this work is development and investigation of algorithmic
and software tools for fast analysis of breast tissue images, detection of tumors
and their classification into two classes: benign or malignant one. This will enable
to provide express analysis of images and raise the quality medical diagnostics.
DATA SET DESCRIPTION
For our investigation we used data set BreaKHis specially created for estimation
efficiency of different approaches and tools for medical images of breast tumor
diagnostics.
Data set BreaKHis [7] contains microscope biopsies from benign and malign
tumors of breast. The images were obtained in clinic research since January 2014
till December 2014.
BreaKHis consists of 7909 clinically representative microscopic images of
breast tumors received from 82 patients with different scale augmentation (40×,
100×, 200×, 400×).
All patients which during this period were investigated in R&D medical lab
with clinical conclusion of breast cancer were invited to take part in this investi-
gation. All data were anonymized. The patterns are generated of biopsy breast
slides colored with hematoxulin and eosin (HE). The patterns are collected by
surgery biopsy prepared for histologic research and marked by pathologist-
anatomists of R&D lab. The main goal is to preserve original structure of tissue
and molecular composition which allows to observe it with optical microscope.
For investigation all images were split into slides of size 3 mkm. The final con-
clusion of each case was made by experienced pathologist-anatomist which was
confirmed by additional investigation such as immune histology-chemistry (IHC).
The microscope system Olympus BX-50 with augmentation 3.3 connected
with digital camera Samsung SCC-131AN, is used for obtaining digitized images
of breast tissue. Images were obtained in 3-channels color space True color (24
bits value, 8 bits color channels RGB) with magnification coefficients 40×, 100×,
200×, and 400×. In the fig. 1–4 four images are presented with four magnification
coefficients: (a) 40 ×, (b) 100 ×, (c) 200 ×, (d) 400 × — obtained from one slide of
breast tumor which contains malignant tumor (breast cancer) Separated rectan-
gular ( added by hand for illustrative aims) — region of interest (ROI) which was
chosen by pathologist-anatomist will be described in the next section. In the fig. 5
the image of benign tumor is presented.
Up to date dataset BreakHis consists of 7909 images, divided into benign
and malignant tumors. Table 1 presents the distribution of images [7].
T a b l e 1 . Distribution of images by magnification coefficients and class
Magnification Benign Malignant Total
40× 625 1370 1995
100× 644 1437 2081
200× 623 1390 2013
400× 588 1232 1820
Total 2480 5429 7909
Number of patients 24 58 82
YU. Zaychenko, G. Hamidov, I. Varga
ISSN 1681–6048 System Research & Information Technologies, 2018, № 4 40
Fig. 1. Slide of malignant tumor with magnification 40×
Fig. 2. Slide of malignant tumor with magnification 100×
Fig. 3. Slide of malignant tumor with magnification 200×
Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network …
Системні дослідження та інформаційні технології, 2018, № 4 41
CONVOLUTIONAL NEURAL NETWORKS: BRIEF DESCRIPTION
A CNN model is a state-of-the-art method that has been largely utilized for image
processing. A CNN model has the ability to extract global features in a hierarchi-
cal manner that ensures local connectivity as well as the weight-sharing property.
It consists of the following layers [9, 10].
Convolutional Layer: The Convolutional layer is considered as the main
working ingredient in a CNN model and plays a vital determining part of this
model. A kernel (filter), which is basically an n×n matrix successively goes
through all the pixels and extracts the information from them.
Stride and Padding: The number of pixels a kernel moves in a step is de-
termined by the stride size; conventionally, the size of the stride is set to 1. Let we
have an input data matrix of size 5×5, which is scanned with a 3×3 kernel. When
Fig. 4. Slide of malignant tumor with magnification 400×
Fig. 5. Slide of benign tumor with magnification 100×
YU. Zaychenko, G. Hamidov, I. Varga
ISSN 1681–6048 System Research & Information Technologies, 2018, № 4 42
we use a 3×3 kernel, and stride size 1, then the convolved output is a 3×3 matrix;
however, when we use stride size 2, the convolved output is 2×2. Interestingly, if
we use a 5×5 kernel on the above input matrix with stride 1, the output will be a
1×1 matrix. Thus, the size of the output image changes with both the size of the
stride and the size of the kernel. To overcome this drawback, we can utilize extra
rows and columns at the end of the matrices that contain 0 s. This adding of
rows and columns that contain only zero values is known as zero padding.
Nonlinear Performance: Each layer of the NN produces linear output, and
by definition adding two linear functions will also produce another linear output.
Due to the linear nature of the output, adding more NN layers will show the same
behavior as a single NN layer. To overcome this issue, a rectifier functions such as
Rectified Linear Unit (ReLU), Leaky ReLU, Tanh, Sigmoid, etc., are introduced
to make the output nonlinear.
Pooling Operation: A CNN model produces a large amount of feature in-
formation. To reduce the feature dimensionality, a down-sampling method named
a pooling operation has been performed. A few pooling operation methods are
well known such as [9,10]: Max Pooling, Average Pooling.
For our analysis, we have utilized the Max Pooling operation that selects the
maximum values within a particular patch.
Drop-Out: Due to the overtraining of the model, it shows very poor per-
formance on the test dataset, which is known as over-fitting. These over-fitting
issues have been controlled by removing some of the neurons from the network,
which is known as Drop-Out.
Decision Layer: For the classification decision, at the end of a CNN model,
a decision layer (usually MLP) is introduced. Normally, a Softmax layer or SVM
layer is introduced for this purpose. This layer contains a normalized exponential
function and calculates the loss function for the data classification.
CNN MODEL FOR IMAGECLASSIFICATION
In the next fig. 6 the architecture of VGG-16 is presented which was used in our
work as detector of informative features. It was trained by different algorithms:
stochastic gradient descent (SCD), differential evolution [14, 15] and basin hop-
ping [11].
As classifier of obtained features in our research it was suggested to use
FNN NEFClass. FNN NEFClass was firstly suggested by D. Nauck and W. Kruse
in [12]. It was modified and developed in [13, 14] (so-called FNN NEFClass M)
The learning algorithms for FNN NEFClass: stochastic gradient SG, conjugate
gradient descent (CGS) and genetic algorithm were developed and investigated in
[14] for the problem of optical images pattern recognition.
FNN NEFCLASS was successfully applied for analysis of medical images
of cervix tissue obtained with use of colposcope and diagnostics [16]. The main
advantages of FNN NEFClass as classifier are: possibility to work with incom-
plete and fuzzy input data; performing fuzzy classification of input patterns
Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network …
Системні дослідження та інформаційні технології, 2018, № 4 43
(images) using so-called membership functions; speed and high accuracy
[13, 14, 15].
EXPERIMENTAL INVESTIGATIONS AND ANALYSIS
As it was already mentioned in our investigation pre-trained CNN VGG-16 was
used. Method of training transfer was applied for this purpose. There are two
main training scenarios:
Features extraction. In this case the last full-connected layer is deleted and
the rest part of CNN is used as extractor for new data sets.
Fine tuning. In this case new data set is used for fine training of previously
pre-trained neural network. In our research CNN VGG-16 was used for features
extraction in medical images of breast tumors. After that the found features were
fed as input data to FNN NEFClass. As algorithms of training FNN three algo-
rithms were used: basin hopping [11], stochastic gradient descent and differential
evolution [15].
EXPERIMENTS DESCRIPTION
The series of experiments were carried out and the results were compared with
works of predecessors. In the following tables 2, 3 the results of classification
with different parameters are presented. All sample was divided into training and
testing subsamples with ratio 80% / 20%.
In the first experiment we varied the number of linguistic variables (terms)
and rules that to determine the best parameters values (table 2).
Fig. 6. Convolutional neural network VGG-16
YU. Zaychenko, G. Hamidov, I. Varga
ISSN 1681–6048 System Research & Information Technologies, 2018, № 4 44
T a b l e 2 . Classification results of FNN NEFClass
Initial number of fuzzy sets
(linguistic terms) /
number of rules
40×, % 100×, % 200×, % 400×, %
2/2 73 74 74,2 73,5
4/2 75,3 74,8 75,7 75,4
6/2 78,2 79 78,4 78
8/2 76 75,4 76,5 75,8
2/4 75 74 73,8 73
4/4 78,3 76,3 75,7 75,4
6/4 82 83 82,4 83,2
8/4 82,2 81,5 81,5 83,8
2/6 75,4 73,8 74,4 73,2
4/6 90 91 90,5 90
6/6 89 89,7 90,2 89,5
8/6 90,3 90,5 92 91,2
4/8 89,3 89,8 89,7 89,3
6/8 89,2 88 89,4 88,4
8/8 88 87,2 87,2 87
From this table one can readily see that beginning from 6 fuzzy sets per vari-
able and 6 rules the accuracy doesn’t increase but complexity of training raises.
As it follows from the table for two classes the best values of parameters for
FNN NEFClass are 4 fuzzy sets per variable and 6 rules. For comparison let’s
present the results of the previous work obtained with different classifiers for the
same problem [6] (see table 3).
Table 3. Comparison of different classifiers accuracy
Classifier/magnification
coefficient
40×,% 100×, % 200×, % 400×, %
Linear SVM 89 89 88 88
Polynomial SVM 88 90 89 85
Random forest 89,18 88 87,74 80
NEFClass 90 91 90,5 90
As we can see from the table 3 FNN NEFClass shows better results than
previous classifiers: SVM machine and Random forest suggested in [6].
In our work for training of FNN NEFClass three algorithms were applied,
namely, basin hopping, stochastic gradient descent and differential evolution. Us-
ing algorithms basin hopping and stochastic gradient descent we obtained ap-
proximately equal results that may mean to be close to optimal results while the
training results of differential evolution appeared to be much worse.
It’s worth to note that in this problem the number of features extracted by
CNN VGG16 was very large — 4096 features. Therefore it was decided to cut the
Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network …
Системні дослідження та інформаційні технології, 2018, № 4 45
number of features and reduce dimensionality of classification problem. For this
aim principal components method (PCM) [17] was applied. In the table 4 the re-
sults of such reduction are presented.
T a b l e 4 . The dependence of total variance on number of components
and approximate training time
Number of principal
components
Variation
Approximate training time
(in hours)
100 0,84058 ~2
200 0,89736 ~3
250 0,91232 ~4
500 0,95486 ~9
From the table 4 it follows that the results of reduction with 250 principal
components are most acceptable as the complexity of training increases approxi-
mately proportional to dimension of input data. Due to lack of time the next ex-
periments were performed using data with magnification factor 100× (2081 im-
ages). In the next table 5 the accuracy of classification is presented with various
NEFClass parameters.
T a b l e 5 . Classification accuracy with 250 features
Number of fuzzy sets / number of rules 100×, %
4/4 80,64
4/6 87,24
4/8 88,18
In the table 6 the dependence of classification accuracy versus number of
features is presented. One can see from this table that accuracy decreased only by
some percent due such features reduction. But by this reduction we substantially
have cut the training time.
Table 6. Classification accuracy with different number of features
Number of linguistic terms,
number of rules/number of features
100 250 4096
4/4 75,23% 80,64% 86,3%
4/6 83,34% 87,24% 91%
4/8 84,21% 88,18% 89,8%
From this table one can easily see that the accuracy drops with decrease of
features number but insignificant by 3–5% if compare results with 100 and 250
features. For comparison the classification with the full set of features 4096 was
performed and we detected that with decrease features number in 20 times the
accuracy falls only by 3–5%, in average. This conclusion confirms the efficiency
of PCM method application for reduction of dimensionality of medical images
classification problems.
YU. Zaychenko, G. Hamidov, I. Varga
ISSN 1681–6048 System Research & Information Technologies, 2018, № 4 46
CONCLUSION
The problem of analysis of breast tissue medical images and classification
of detected tumor in two classes: benign and malignant is considered and dis-
cussed.
For pattern recognition of breast tumors new hybrid CNN- FNN network
is suggested in which CNN VGG 16 is used for informative features extraction
while FNN NEFClass is used for classification of detected tumors.
For training FNN NEFClass algorithms basin hopping, stochastic gradient
descent and differential evolution were suggested and their efficiency investi-
gated.
The experimental investigations of suggested hybrid CNN-FNN network
in the problem of classification real images of breast tumors using dataset
BreakHis were carried out.
The comparison of classification accuracy of the suggested hybrid CNN-
FNN network with known work based on use of classification algorithms SVM
and Random forest was performed which confirmed the efficiency of the sug-
gested approach.
The problem of reducing number of features in medical images classifi-
cation problem using PCM method was investigated and its efficiency explored.
REFERENCES
1. Boyle P. World Cancer Report 2012 / P. Boyle, B. Levin, Eds. — Lyon: IARC,
2012. — Available at: http://www.iarc.fr/en/publications/pdfsonline/wcr/2008/
wcr_2012.pdf
2. Lakhani S.R. WHO classification of tumours of the breast / S.R. Lakhani, S. Schnitt
et al. — 4th ed. — Lyon: WHO Press, 2012.
3. Zhang Y. Breast cancer diagnosis from biopsy images with highly reliable random
subspace classifier ensembles / Y. Zhang, B. Zhang, F. Coenen et al. // Machine
Vision and Applications. — 2013. — Vol. 24, N. 7. — P. 1405– 1420.
4. Zhang Y. One-class kernel subspace ensemble for medical image classification /
Y. Zhang, B. Zhang, F. Coenen et al. // EURASIP Journal on Advances in Signal
Processing. — 2014. — Vol. 2014, N 17. — P. 1–13.
5. Doyle S. Automated grading of breast cancer histopathology using spectral clustering
with textural and architectural image features / S. Doyle, S. Agner, A. Madab-
hushi et al. // in Proceedings of the 5th IEEE International Symposium on
Biomedical Imaging (ISBI): From Nano to Macro. — Vol. 61. — IEEE, May
2008. — P. 496–499.
6. Singh Aditi. Classifying Biological Images Using Pre-trained CNNs / Aditi Singh,
Hadi Mansourifar, Hasnain Bilgrami et al. — Available at:
https://docs.google.com/document/d/1H7xVK7nwXcv11CYh7hl5F6pM0m218F
QloAXQODP-Hsg/edit?usp=sharing
7. Spanhol F. A dataset for breast cancer histopathological image classification /
F. Spanhol, L.S. Oliveira, C. Petitjean et al. // IEEE Transactions of Biomedical
Engineering, 2016.
8. Bengio Y. Representation learning: A review and new perspectives / Y. Bengio,
A. Courville, P. Vincent // IEEE Transactions on Pattern Analysis and Machine
Intelligence. — 2013. — Vol. 35. — P. 1798–1828.
Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network …
Системні дослідження та інформаційні технології, 2018, № 4 47
9. LeCun Y. Deep learning / Y. LeCun, Y. Bengio, G. Hinton // Nature. — 2015. —Vol.
521. — P. 436–444.
10. Krizhevsky A. Imagenet classification with deep convolutional neural networks /
A. Krizhevsky, I. Sutskever, G.E. Hinton // Advances in Neural Information Pro-
cessing Systems. — 2012. — Vol. 25. — P.1097–1105.
11. Olson B. Basin Hopping as a General and Versatile Optimization Framework for the
Characterization of Biological Macromolecules / B. Olson, I. Hashmi, K. Molloy
et al. // Advances in Artificial Intelligence. —2012. — Article ID 674832.
12. Nauck Detlef. New learning strategies for NEFCLASS / Detlef Nauck, Rudolf Kruse
// In Proc. Seventh International Fuzzy Systems Association World Congress
IFSA’97. — Prague: Academia Prague, 1997. — Vol. IV. — P. 50–55.
13. Zaychenko Yu.P. Fuzzy neural networks for economic data classification /
Yu.P. Zaychenko, Fatma Sevaee, A.V. Matsak // Vestnik of National Technical
University of Ukraine “KPI”, section “Informatic, control and computer engi-
neering”. — 2004. — Vol. 42. — P. 121–133.
14. Zaychenko Yu.P. The investigations of fuzzy neural networks in the problems
of electro-optical images recognition / Yu.P. Zaychenko, I.M. Petrosyuk,
M.S. Jaroshenko // System research and information technologies. — 2009. —
N 4. — P. 61–76.
15. Zgurovsky M. The Fundamentals of Computational Intelligence: System Approach /
M. Zgurovsky, Yu. Zaychenko // Switzerland: Springer International Publishing
AG. — 2016. — 308 p.
16. Zaychenko Yu. Recognition of objects on Optical Images in Medical Diagnostics Us-
ing Fuzzy Neural Network NEFClass / Yu, Zaychenko, V. Huskova // Intern.
Journal Information Models and Analysis. — 2015. — Vol. 4, N 1. — P. 13–22.
17. Jindal N. Enhanced Face Recognition Algorithm using PCA with Artificial Neural
Networks / N. Jindal, V. Kumar // International Journal of Advanced Research in
Computer Science and Software Engineering. — 2013. — Vol. 3. — P. 864–872.
Received 27.08.2018
From the Editorial Board: the article corresponds completely to submitted
manuscript.
|
| id | journaliasakpiua-article-152060 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:24:15Z |
| publishDate | 2018 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/fb/d3ea0b02d082f35237e2c063a3e4a8fb.pdf |
| spelling | journaliasakpiua-article-1520602019-04-26T15:57:21Z Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network Диагностика медицинских изображений опухолей с применением гибридных сверточных нечетких нейронных сетей Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж Zaychenko, Yuriy Hamidov, G. Varga, I. medical diagnostics breast cancer classification FNN CNN hybrid network dimensionality reduction PCM медицинская диагностика рак молочной железы гибридная сверточная нейросеть редукция размерности медична діагностика рак молочної залози гібридна згорткова нейромережа редукція вимірності The problem of classification of breast tumors on medical images is con-sidered. For its solution the new class of convolutional neural networks-hybrid CNN–FNN network is developed in which convolutional neural network VGG-16 is used as the feature extractor while fuzzy neural network NEFClass is used as the classifier. Training algorithms of FNN were implemented. The experimental investigations of the suggested hybrid network on the standard data set were carried out and comparison with known results was performed. The problem of data dimensionality reduction is considered and application of PCM method is investigated. Рассмотрена проблема классификации опухолей молочной железы по медицинским изображениям. Для ее решения предложен новый класс сверточных сетей — гибридная нечеткая сверточная нейронная сеть, в которой сверточная нейронная сеть VGG‑16 используется как экстрактор признаков изображений, а нечеткая нейронная сеть NEFClass — как классификатор. Разработаны и исследованы алгоритмы обучения гибридной нейронной сети. Проведены экспериментальные исследования предложенной гибридной сверточной нечеткой нейронной сети на стандартной базе данных Breakhis и выполнено сравнение с известными результатами, что позволило оценить ее эффективность. Рассмотрена проблема уменьшения размерности задачи классификации и для ее решения предложен и исследован метод главных компонент. Розглянуто проблему класифікації пухлин молочної залози за медичними зображеннями. Для її вирішення запропоновано новий клас згорткових мереж — гібридну нечітку згорткову нейронну мережу, в якій згорткова мережа VGG-16 використовується як екстрактор ознак зображення, а нечітка нейронна мережа NEFClass — як класифікатор. Розроблено та досліджено алгоритми навчання гібридної згорткової мережі. Проведено експериментальні дослідження запропонованої гібридної згорткової мережі на стандартній базі даних Breakhis та виконано порівняння з відомими результатами, що дозволило оцінити її ефективність. Розглянуто проблему зменшення вимірності задачі класифікації і для її вирішення запропоновано та досліджено метод головних компонент. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2018-12-18 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/152060 10.20535/SRIT.2308-8893.2018.4.03 System research and information technologies; No. 4 (2018); 37-47 Системные исследования и информационные технологии; № 4 (2018); 37-47 Системні дослідження та інформаційні технології; № 4 (2018); 37-47 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/152060/151390 Copyright (c) 2021 System research and information technologies |
| spellingShingle | медична діагностика рак молочної залози гібридна згорткова нейромережа редукція вимірності Zaychenko, Yuriy Hamidov, G. Varga, I. Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| title | Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| title_alt | Medical images of breast tumors diagnostics with application of hybrid CNN–FNN network Диагностика медицинских изображений опухолей с применением гибридных сверточных нечетких нейронных сетей |
| title_full | Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| title_fullStr | Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| title_full_unstemmed | Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| title_short | Діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| title_sort | діагностика медичних зображень пухлин з використанням гібридних нечітких згорткових нейронних мереж |
| topic | медична діагностика рак молочної залози гібридна згорткова нейромережа редукція вимірності |
| topic_facet | medical diagnostics breast cancer classification FNN CNN hybrid network dimensionality reduction PCM медицинская диагностика рак молочной железы гибридная сверточная нейросеть редукция размерности медична діагностика рак молочної залози гібридна згорткова нейромережа редукція вимірності |
| url | https://journal.iasa.kpi.ua/article/view/152060 |
| work_keys_str_mv | AT zaychenkoyuriy medicalimagesofbreasttumorsdiagnosticswithapplicationofhybridcnnfnnnetwork AT hamidovg medicalimagesofbreasttumorsdiagnosticswithapplicationofhybridcnnfnnnetwork AT vargai medicalimagesofbreasttumorsdiagnosticswithapplicationofhybridcnnfnnnetwork AT zaychenkoyuriy diagnostikamedicinskihizobraženijopuholejsprimeneniemgibridnyhsvertočnyhnečetkihnejronnyhsetej AT hamidovg diagnostikamedicinskihizobraženijopuholejsprimeneniemgibridnyhsvertočnyhnečetkihnejronnyhsetej AT vargai diagnostikamedicinskihizobraženijopuholejsprimeneniemgibridnyhsvertočnyhnečetkihnejronnyhsetej AT zaychenkoyuriy díagnostikamedičnihzobraženʹpuhlinzvikoristannâmgíbridnihnečítkihzgortkovihnejronnihmerež AT hamidovg díagnostikamedičnihzobraženʹpuhlinzvikoristannâmgíbridnihnečítkihzgortkovihnejronnihmerež AT vargai díagnostikamedičnihzobraženʹpuhlinzvikoristannâmgíbridnihnečítkihzgortkovihnejronnihmerež |