Використання згорткових нейронних мереж для діагностування раку молочної залози
During the last few years, Convolutional Neural Networks (CNN) have been widely used in Computer-Aided Detection and the medical image analysis. The main idea of this paper is to modify CNN’s architectures to achieve the better sensitivity and the precision for detecting breast cancer at an early st...
Saved in:
| Date: | 2019 |
|---|---|
| Main Authors: | , , |
| Format: | Article |
| Language: | English |
| Published: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2019
|
| Subjects: | |
| Online Access: | https://journal.iasa.kpi.ua/article/view/183548 |
| 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_ | 1867334398902272000 |
|---|---|
| author | Naderan, Maryam Zaychenko, Yuriy Napoli, Amedeo |
| author_facet | Naderan, Maryam Zaychenko, Yuriy Napoli, Amedeo |
| author_institution_txt_mv | [
{
"author": "Maryam Naderan",
"institution": "Educational and Scientific Complex \"Institute for Applied System Analysis\" of the National Technical University of Ukraine \"Igor Sikorsky Kyiv Polytechnic Institute\", Kyiv"
},
{
"author": "Yuriy Zaychenko",
"institution": "Educational and Scientific Complex \"Institute for Applied System Analysis\" of the National Technical University of Ukraine \"Igor Sikorsky Kyiv Polytechnic Institute\", Kyiv"
},
{
"author": "Amedeo Napoli",
"institution": "Centre De Recherche Inria Nancy Grand-Est"
}
] |
| author_sort | Naderan, Maryam |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2020-03-02T17:05:10Z |
| description | During the last few years, Convolutional Neural Networks (CNN) have been widely used in Computer-Aided Detection and the medical image analysis. The main idea of this paper is to modify CNN’s architectures to achieve the better sensitivity and the precision for detecting breast cancer at an early stage compared to existing methods. For this purpose, several factors were considered before CNN training such as the data processing, model, dataset, etc. In the proposed model the following hyperparameters were the following: the dropout rate 0,2, epoch 38 and batch size 33. Besides the hyperparameters, two fully connected layers in the modified model were used. An average recall (sensitivity) in the recent works was 74%. The precision and recall of proposed model for breast cancer classification were 66,66% and 85,7%, respectively. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2019.4.09 |
| first_indexed | 2025-07-17T10:26:27Z |
| format | Article |
| fulltext |
M. Naderan, Y. Zaychenko, A. Napoli, 2019
Системні дослідження та інформаційні технології, 2019, № 4 85
UDC 004.855.5
DOI: 10.20535/SRIT.2308-8893.2019.4.09
USING CONVOLUTIONAL NEURAL NETWORKS
FOR BREAST CANCER DIAGNOSING
M. NADERAN, Yu. ZAYCHENKO, A. NAPOLI
Abstract. During the last few years, Convolutional Neural Networks (CNN) have
been widely used in Computer-Aided Detection and the medical image analysis. The
main idea of this paper is to modify CNN’s architectures to achieve the better sensi-
tivity and the precision for detecting breast cancer at an early stage compared to ex-
isting methods. For this purpose, several factors were considered before CNN train-
ing such as the data processing, model, dataset, etc. In the proposed model the
following hyperparameters were the following: the dropout rate 0,2, epoch 38 and
batch size 33. Besides the hyperparameters, two fully connected layers in the modi-
fied model were used. An average recall (sensitivity) in the recent works was 74%.
The precision and recall of proposed model for breast cancer classification were
66,66% and 85,7%, respectively.
Keywords: convolutional neural networks, deep learning, computer-aided detection,
breast cancer diagnosis, classification.
INTRODUCTION
Breast cancer is one of the most invasive cancers between women. The number of
patients who have this type of cancer is increasing not only in poor countries but
also in developed countries. Based on World Health Organization (WHO) [1],
mammography is cost-effective for analyzing presence of breast cancer in pa-
tients. In other words, mammography is expensive but is the only method for
screening breast cancer that has proven effective. In [2], the authors indicated that
digital mammography (DM) produces the best medical image and is highly rec-
ommended for computer aided detection. In this paper, mammography scans were
used for experiments, since they are of high quality.
The object of this study is mammography screening and the subject of this
study is deep learning methods for diagnosing breast cancer.
In this paper, all methods for diagnosing breast cancer were implemented in
jupyter notebook using keras libraries [3]. We consider recall and precision as
evaluation metrics, since for detecting breast cancer, accuracy cannot be the only
indicator for making decision.
NOVELTY
In this work, we proposed modified Inception V3 model to gain better results
comparing to previous works. In order to that, the best values of hyperparameters
that are playing an important role in the model were chosen. Also, before fully
connected layer dropout with 0.2 rate was added for making model more inde-
M. Naderan, Yu. Zaychenko, A. Napoli
ISSN 1681–6048 System Research & Information Technologies, 2019, № 4 86
pendent to training data. Besides that, two fully connected layers were used in the
modified model to achieve the goal.
RELATED WORKS
Systematic review (SR) has been done in [2]. In this paper authors illustrated how
choosing the right medical images is important. Digital Mammography (DM) is
the most important technique in medicine screening. DM helps to detect tumor
before it develops further. Moreover, authors compare different methods that
were used from 2011–2017 and conclude that SVM has the best result based on
[4, 5, 6, 7].
Currently, there are few works that are considering using convolutional neu-
ral networks for the task. In [8], authors compare machine learning methods, such
as Support Vector Machine (SVM), Decision Tree (C4.5), Naive Bayes (NB) and
k-Nearest Neighbors (k-NN) on white blood cell (WBC) datasets. Based on their
experiments, SVM has the best accuracy with 97,13%. However, for cancer diag-
nosis task precision and recall should be considered as evaluation metrics. In [9],
a modified CNN architecture with nine layers was proposed. From these layers,
six of them were convolution and pooling layers and three were fully connected
layers. The result of this work shows 69,99% and 81,44% accuracy for cancer
classification and necrosis detection respectively.
Authors in [10] consider factors, such as true positive, false positive, preci-
sion and recall for classification using HPBCR (hybrid predictor of breast cancer
recurrence). Sensitivity of the performance in [10] was 77% and the accuracy was
95%, and the same values for SVM and decision tree were 67%, 78% and 75%,
77% respectively.
Authors in [11] propose an [“end-to-end” approach in which a model to clas-
sify local image patches is pre-trained using a fully annotated dataset with region
of interest (ROI)information].The experiment results show ResNet 50 has better
accuracy (97%) comparing to VGG(84%).End-to-end learning process is a type of
Deep learning process when all the parameters are trained jointly, rather than step
by step. The weakness of [11] is that authors are using simple validation. Simple
validation means that the training dataset is split in training and validation sets
such as 70% of the data is used for training and 30% for validation. Instead
K-fold cross validation divides the data into K number of sections/folds where
each fold is used as a testing set at some point. Using K-fold cross validation
helps to prevent overfitting without losing any data [12, 13].
CONVOLUTIONAL NEURAL NETWORK
A Neural Network (NN) is a network of neurons that are used to process informa-
tion. A simple NN includes three layers: input, hidden and output. A Convolu-
tional Neural Network (CNN) is a Deep Learning network which can take in an
input image, assign importance (learnable weights and biases) to various as-
pects/objects in the image and be able to differentiate one from the other. Convo-
lutional neural network has three main layers: Convolutional layer, Pooling layer
and Fully Connected layer. Fig. 1 illustrates the architecture of CNN. The main
Using convolutional neural networks for breast cancer diagnosing
Системні дослідження та інформаційні технології, 2019, № 4 87
difference between Convolutional Neural Network CNN and Neural Network
(NN), is a convolutional part. CNN has extra Convolutional and Pooling layers.
Basically, a fully connected (FC) layer is simple NN. The number of convolu-
tional and pooling layers depend on the model which is used. For example, in [9]
there is no pooling layer but six convolutional layers.
There are three main operation illustrated in fig. 1:
1. Convolution and Non-Linearity (ReLU).
2. Pooling or Sub Sampling.
3. Classification (Fully Connected Layer).
These operations are the main building blocks of each convolutional neural
network. The first three operations are used for features extraction and the outputs
of convolutional part (convolutional, ReLU and pooling) are used as input to the
fully connected layers where classification happens.
The convolutional layer and ReLU function
The main idea of convolutional layer is features extraction. At this stage filters are
applied to the input image for features extraction. In order to this, the filter slides
(orange matrix is called “kernel”) over the image (green matrix) by 1 pixel
(stride) for every position, element wise multiplication is computed (between
the two matrices) and outputs are added in order to get the final integer that
forms a single element of the output matrix (pink matrix), (fig. 2). For exam-
ple, the number “4”, the first pixel on left side, calculated as
By applying
one filter to the image, we will get the first feature. By applying several filters to
the same image, we will get several features that constitute a feature map [14].
The ReLU operation (also called activation function) is performed right after
convolution. At this stage, all negative pixels are changed to zero in order to in-
troduce non-linearity in ConvNet, since most of the real world data would be non-
linear (Convolution is a linear operation – element wise matrix multiplication and
addition, so we account for non-linearity by introducing a non-linear function like
ReLU) [23]. Operation ReLUis calculated using the formula below:
),0(max XY .
Convolution
+ReLU
Poolling Convolution
+ReLU
Poolling Fully
Connected
Fully
Connected
Output
Predictions
2)(
2
1
outputgetartEtotal
Feature Extraction from Image Classificatic
1
2
3
4
Fig. 1. A simple ConvNet; 1 — Dog (0), 2 —Cat (0), 3 — Boat (1), Bird — (0)
M. Naderan, Yu. Zaychenko, A. Napoli
ISSN 1681–6048 System Research & Information Technologies, 2019, № 4 88
The pooling layer
Pooling layer (also called subsampling), (fig. 3) changes the dimensionality of the
feature maps. The input for this layer is the output of convolution (feature maps)
and output is compressed version of the feature maps. Pooling can be calculated
by Max, Average or Sum operation. The example of max pooling is illustrated in
figure 2 where filter 2×2 with stripe 2 is applied to one feature. In this case, we
slide filter over the feature map with stride equal 2. For instance, in order to cal-
culate the first pixel of max pooling matrix in this example “6”, the follow opera-
tion will be considered: 6)6,5,1,1(Max
The fully connected layer
The Fully Connected layer is a traditional Multi-Layer Perceptron that uses Soft-
max activation function in the output layer. This function calculates the probabili-
ties of each target class for the given input. The output of the convolutional and
pooling layers represents high-level features of the input image. The purpose of
the fully connected layer is to use these features for classifying the input image
into various classes based on the training dataset.
METHODS
There are various machine learning algorithms and methods that could be used for
diagnosing breast cancer with mammography, such as Support Vector Machine,
Random Forest, K-Nearest Neighbor, etc. Even though support vector machines
Fig. 2. Convolutional operation on one input image
Fig. 3. Max pooling operation on one of the features
max pool with 22
filters and stride 2
Max (1,1,5,6)=6
x
y
Using convolutional neural networks for breast cancer diagnosing
Системні дослідження та інформаційні технології, 2019, № 4 89
are widely used for different tasks, and have shown good results, convolutional
neural networks are showing better performance among computer vision algo-
rithms because of the ability to extract important features. After each convolu-
tional operation, some features will be extracted and passed to the deeper layers,
these features will become more specific and unique for the input image. For in-
stance, for detecting a face in the picture, at first layer lines and curves of nose are
extracted. In deeper layers, these lines and curves will be used as one feature
(nose) of the face. By doing the same process, features like eyes, ears, lips etc. are
extracted from the input image.
In [15] three different CNN architectures Cifar Net, AlexNet and GoogLe-
Net (Inception) were compared. It requires 5,37 GB of memory and 2h49m to
apply GoogleNet. While for applying CifarNet, it takes 2,25 GB and 7m16s.
Some of the existing CNN models are very well realized. In this case to im-
prove the precision and recall for our task, we can fine tune pre-trained models.
Because, training model from scratch sometimes does not give the expected re-
sults. Since, training network from the scratch requires so many training datasets.
Unfortunately, it is challenging to get access to the vast mammography screens.
Transfer learning is a machine learning technique that helps reuse previous mod-
els for a running task and improve the used model architecture to reach higher
accuracy and f-score.
EXPERIMENT
For conducting experiment Jupiter notebook was used. The machine where the
program was run had an Intel Core i7 processor and NVIDIA GeForce driver.
In this paper two open-source datasets BreaKHis, Breast histology and
Kaggle [16,17,18] were used during the experiment. In BreaKHis dataset, 1271
images were used for training and 70 images were used for testing, in Breast his-
tology set, 200 images are used for training and 37 are used for testing. The data-
set in Kaggle is numerical and 500 data were used for training and 70 data were
used for testing.
Results show that small dataset could affect the accuracy and CNNs cannot
be trained well leading to overfitting or underfitting. Chart 1 (fig. 4) shows accu-
racy comparison of Inception V3 for BreaKHis and Breast histology datasets.
Fig. 4. Comparing accuracy for training and validation sets
Accuracy
Breasthistology BreaKHis
Training Accuracy Validation Accuracy
M. Naderan, Yu. Zaychenko, A. Napoli
ISSN 1681–6048 System Research & Information Technologies, 2019, № 4 90
According to the chart 1, accuracy for validation set in Breast histology data-
set is noticeably lower value than in BreaKHis dataset. In this case our model is
over fitted. This means that the model is doing well on the training set but not for
those data which were never seen before (validation set). Thus, using a large data-
set prevents model from overfitting during training and fine tuning.
Data preprocessing
Data preprocessing is an integral step in Machine Learning as the quality of data
and the useful information that is derived from it directly affects the ability of the
model. Also, it is used to transform the raw data in a useful and efficient format.
Therefore, it is extremely important that the data are preprocessed before feeding
it into the model [19]. Basically, we preprocess the raw data by importing librar-
ies, read data, checking for missing values, checking for categorical data etc. Im-
age Data Generator module [20] was used for augmentation data and make data-
set bigger by creating different versions of one image.
After preprocessing data, we need to separate data into training and valida-
tion sets. Basically, validation set helps model to be well trained for data which
were never seen before. There are two ways to separate the date into training and
validation sets: simple validation and K-fold cross validation. Simple validation
splits data into two sets where one part of data is used for training and another
part is used for validation, whereas, in K-fold cross validation, dataset is split into
K fold (part). At the first iteration, the first fold is used for validation set and
(K–1) folds are used for training set. At the second iteration, the second fold is
used for validation and the rest of the folds are used for training. This process is
repeated until each fold has been used as validation set.
In Table 1, sensitivity (Recall) with K-fold validation is higher than with
simple validation, 70% and 31,3% respectively. When simple validation is used
during the experiment, some data could be missed in case when some images are
not considered. As a result, CNN won’t be trained well or over fitted.
T a b l e 1 . Comparing simple validation and K-fold validation for Inception V3
Simple validation K-fold validation
Precision Recall F1-score Precision Recall F1-score
30,59% 31,3% 30,11% 65,7% 70% 67,78%
Table 2 shows the difference between our modified model and pre-trained
model. Modified model is using the best weight to achieve better results, and
based on Table 3, precision and recall (sensitivity) in the modified model are
66,66% and 85,7% respectively. In cancer diagnosing, the more important factor
is sensitivity, which can be calculated as (6,1). In other word, we should avoid
misclassification in case when the actual class is yes (cancer) and model predic-
tion is no (no cancer).
T a b l e 2 . Comparing fine tune and pre-trained model (Inception V3)
Model Accuracy, % Precision, % Recall, % F1-Score, %
Inception V3 79,31 65,7 70 67,78
Modified Model 78,59 66,66 85,7 74,99
Using convolutional neural networks for breast cancer diagnosing
Системні дослідження та інформаційні технології, 2019, № 4 91
Based on Table 3, sensitivity (recall), precision and f1-score are calculated
as following:
%7,85
14
12
yesActual
TP
ySensitivit ;
%66,66
18
12
yesredictedP
TP
recisionP ;
%99,745*
recisionPySensitivit
recisionPySensitivit
ySensitivit .
T a b l e 3 . Confusion matrix
Index Predicted No Predicted Yes
Actual No TN = 9 FP = 5
Actual Yes FN = 1 TP = 12
Table 4 shows the comparison of machine learning algorithms on Kaggle dataset.
T a b l e 4 . Comparing machine learning algorithms on Kaggle dataset
ML algorithms Accuracy F1_score Recall Precision
LR 0,964912281 0,953846154 0,98412698 0,925373134
KNN 0,947368421 0,929133858 0,93650794 0,921875
SVM 0,959064327 0,945736434 0,96825397 0,924242424
NB 0,923976608 0,897637795 0,9047619 0,890625
DT 0,935672515 0,916030534 0,95238095 0,882352941
RF 0,964912281 0,950819672 0,92063492 0,983050847
Based on results in Table 4, Logistic Regression (LR) has shown better re-
sults comparing to other methods. Using LR method, we have achieved F1_Score
and Recall 95,38% and 98,41% respectively. In future work, we will use CNN in
numerical dataset and then will compare the obtained results with LR.
CONCLUSION AND DISCUSSION
Convolutional neural networks (CNN) is a special architecture of artificial neural
networks. One of the most popular uses of this architecture is image classifica-
tion. The reason is that it starts from lower abstraction and go deeper into higher
abstraction. Convolutional neural network first starts to define curve and line in
an image at higher layers and extracts general feature maps. While going deeper
in lower (deeper) layers, the feature maps become more and more specific and
unique.
The modified model has dropout rate = 0,2, epochs = 38 and batch size = 33.
Besides choosing the right hyperparameters, in the model two fully connected
M. Naderan, Yu. Zaychenko, A. Napoli
ISSN 1681–6048 System Research & Information Technologies, 2019, № 4 92
layers were used. The model shows better result comparing the pre-trained model
(Inception V3) with sensitivity of 85,7%.
Besides choosing the right model, the quality of the image is selected has an
essential role on the result. In this paper, Digital Mammography (DM) is used,
since the quality of scans in mammography is better than other medical scans.
Even though DM is expensive but is the only effective method for screening
breast cancer that has proven effective. Moreover, it helps to detect tumors at a
very early stage.
In future, we plan to improve the architecture of the current model to reach
better results.
REFERENCE
1. Breast cancer: prevention and control. — Available at: https://www.who.
int/cancer/detection/breastcancer/en/
2. Yassin N. Machine learning techniques for breast cancer computer aided diagnosis
using different image modalities: A systematic review / N. Yassin, S. Omran, E.
El Houby, H. Allam // Computer Methods and Programs in Biomedicine. —
2018. — Vol. 156. — P. 25–45.
3. Keras: The Python Deep Learning library. — Available at: https://keras.io/
4. Sharma S. Computer-aided diagnosis of malignant mammograms using zernike mo-
ments and SVM / S. Sharma, P. Khanna // Journal of Digital Imaging. —
2015. — 28 (1). — P. 77–90.
5. Beheshti S.M.A. An efficient fractal method for detection and diagnosis of breast
masses in mammograms / S.M.A. Beheshti et al. // Journal of Digital Imaging. —
2014. — 27 (5). — P. 661–669.
6. Azar A.T. Decision tree classifiers for automated medical diagnosis / A.T. Azar,
S.M. El-Metwally // Neural Comput. Appl. — 2013. — 23 (7). — P. 2387–2403.
7. Jian W. Computer-aided diagnosis of breast microcalcifications based on dual-tree
complex wavelet transform / W. Jian, X. Sun, S. Luo // Biomed. Eng. Online. —
2012. — 11 (1). — p. 96.
8. Hiba A. Using Machine Learning Algorithms for Breast Cancer Risk Prediction and
Diagnosis / A. Hiba, M. Hajar, M. Hassan, N. Thomas // The 6th International
Symposium on Frontiers in Ambient and Mobile Systems. Procedia Computer
Science 83. — 2016. — P. 1064–1069.
9. Sharmaa H. Deep convolutional neural networks for automatic classification of gas-
tric carcinoma using whole slide images in digital histopathology / H. Sharmaa,
N. Zerbe, I. Klempert et al. // Computerized Medical Imaging and Graphics.—
2017. — 61. — P. 2–13.
10. Mohammad M. A Hybrid Computer-aided-diagnosis System for Prediction of Breast
Cancer Recurrence (HPBCR) Using Optimized Ensemble Learning /
M. Mohammad, M. Hamid, M. Marjan et al. // Computational and Structural Bio-
technology Journal. — 2017. — 15. — P. 75–85.
11. Shen Li. Deep Learning to Improve Breast Cancer Early Detection on Screening
Mammography / Li Shen, Laurie R. Margolies, Joseph H. Rothstein et al. Pub-
lished online Mar 15. — doi: 10.1038/s41598-018-22437-z. 2018.
12. K-Fold Cross Validation. — Available at: https://medium.com/ datadriveninvestor/
K-fold -cross-validation-6b8518070833
13. Rakhlin A. Deep Convolutional Neural Networks for Breast Cancer Histology Image
Analysis / A. Rakhlin, A. Shvets, V. Iglovikov. — ICIAR 2018 Grand Challenge.
arXiv:1802.00752v2. 2018.
Using convolutional neural networks for breast cancer diagnosing
Системні дослідження та інформаційні технології, 2019, № 4 93
14. An Intuitive Explanation of Convolutional Neural Networks. — Available at:
https://ujjwalkarn.me/2016/08/11/intuitive-explanation-convnets/
15. Shin Hoo-Chang. Deep Convolutional Neural Networks for Computer-Aided Detec-
tion: CNN Architectures, Dataset Characteristics and Transfer Learning / Hoo-
Chang Shin, Holger R. Roth, Mingchen Gao et al. // IEEE Transactions on medi-
cal imaging. — 2016. — Vol. 35, N 5.
16. Breast cancer dataset from breakhis. — Available at: https://www.kaggle.
com/kritika397/breast-cancer-dataset-from-breakhis
17. Bioimaging Challenge 2015 Breast Histology Dataset. — Available at:
https://rdm.inesctec.pt/dataset/nis-2017-003
18. Breast histopathology. — Available at: https://www.kaggle.com/ paultimothy-
mooney/breast-histopathology-images#IDC_regular_ps50_idx5.zip
19. Sharmaa H. Deep convolutional neural networks for automatic classification of gas-
tric carcinoma using whole slide images in digital histopathology / H. Sharmaa,
N. Zerbe, I. Klempert // Computerized Medical Imaging and Graphics. —
2017. — 61. — P. 2–13.
20. Image Preprocessing. — Available at: https://keras.io/preprocessing/image/
Received 21.11.2019
____________________________
From the Editorial Board: the article corresponds completely to submitted manuscript.
|
| id | journaliasakpiua-article-183548 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:26:27Z |
| publishDate | 2019 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/9f/2185e89444f6543f642c204ec9bbc69f.pdf |
| spelling | journaliasakpiua-article-1835482020-03-02T17:05:10Z Using convolutional neural networks for breast cancer diagnosing Использование сверточных нейронных сетей для диагностики рака молочной железы Використання згорткових нейронних мереж для діагностування раку молочної залози Naderan, Maryam Zaychenko, Yuriy Napoli, Amedeo convolutional neural networks deep learning computer-aided detection breast cancer diagnosis classification згорткові нейронні мережі глибоке навчання комп'ютерне виявлення діагностика раку молочної залози класифікація сверточные нейронные сети глубокое обучение компьютерное обнаружение диагностика рака молочной железы классификация During the last few years, Convolutional Neural Networks (CNN) have been widely used in Computer-Aided Detection and the medical image analysis. The main idea of this paper is to modify CNN’s architectures to achieve the better sensitivity and the precision for detecting breast cancer at an early stage compared to existing methods. For this purpose, several factors were considered before CNN training such as the data processing, model, dataset, etc. In the proposed model the following hyperparameters were the following: the dropout rate 0,2, epoch 38 and batch size 33. Besides the hyperparameters, two fully connected layers in the modified model were used. An average recall (sensitivity) in the recent works was 74%. The precision and recall of proposed model for breast cancer classification were 66,66% and 85,7%, respectively. В течение последних нескольких лет сверточные нейронные сети широко используются в компьютерной диагностике и анализе медицинских изображений. Основная идея работы состояла в модифицировании архитектуры CNN для достижения большей чувствительности и точности в целях выявления рака молочной железы на ранних стадиях по сравнению с уже существующими методами. Для этого перед обучением CNN рассмотрено несколько факторов, таких как предварительная обработка данных, модель, набор данных и др. В предложенной модели использовались гиперпараметры dropout rate 0,2, epoch 38 и batch size 33, а также два полносвязанных слоя в модифицированной модели. Средний показатель полноты (чувствительности) в последних работах составляет 74%. Точность и полнота предлагаемой модели классификации рака молочной железы составили 66,66% и 85,7% соответственно. Протягом останніх кількох років згорткові нейронні мережі широко використовуються в комп’ютерній діагностиці та аналізі медичних зображень. Основна ідея роботи полягала в розробленні модифікованої архітектури CNN для досягнення більшої чутливості і точності для виявлення раку молочної залози на ранніх стадіях порівняно з уже існуючими методами. Для цього перед навчанням CNN розглянуто декілька факторів, таких як попереднє оброблення даних, модель, набір даних і т.ін. У запропонованій моделі використовувалися гіперпараметри dropout rate 0,2, epoch 38 і batch size 33, а також два повнозв’язні шари в модифікованій моделі. Середній показник повноти (чутливості) в останніх працях становить 74%. Точність і повнота запропонованої моделі класифікації раку молочної залози склала 66,66% і 85,7% відповідно. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2019-12-23 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/183548 10.20535/SRIT.2308-8893.2019.4.09 System research and information technologies; No. 4 (2019); 85-93 Системные исследования и информационные технологии; № 4 (2019); 85-93 Системні дослідження та інформаційні технології; № 4 (2019); 85-93 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/183548/190141 Copyright (c) 2021 System research and information technologies |
| spellingShingle | згорткові нейронні мережі глибоке навчання комп'ютерне виявлення діагностика раку молочної залози класифікація Naderan, Maryam Zaychenko, Yuriy Napoli, Amedeo Використання згорткових нейронних мереж для діагностування раку молочної залози |
| title | Використання згорткових нейронних мереж для діагностування раку молочної залози |
| title_alt | Using convolutional neural networks for breast cancer diagnosing Использование сверточных нейронных сетей для диагностики рака молочной железы |
| title_full | Використання згорткових нейронних мереж для діагностування раку молочної залози |
| title_fullStr | Використання згорткових нейронних мереж для діагностування раку молочної залози |
| title_full_unstemmed | Використання згорткових нейронних мереж для діагностування раку молочної залози |
| title_short | Використання згорткових нейронних мереж для діагностування раку молочної залози |
| title_sort | використання згорткових нейронних мереж для діагностування раку молочної залози |
| topic | згорткові нейронні мережі глибоке навчання комп'ютерне виявлення діагностика раку молочної залози класифікація |
| topic_facet | convolutional neural networks deep learning computer-aided detection breast cancer diagnosis classification згорткові нейронні мережі глибоке навчання комп'ютерне виявлення діагностика раку молочної залози класифікація сверточные нейронные сети глубокое обучение компьютерное обнаружение диагностика рака молочной железы классификация |
| url | https://journal.iasa.kpi.ua/article/view/183548 |
| work_keys_str_mv | AT naderanmaryam usingconvolutionalneuralnetworksforbreastcancerdiagnosing AT zaychenkoyuriy usingconvolutionalneuralnetworksforbreastcancerdiagnosing AT napoliamedeo usingconvolutionalneuralnetworksforbreastcancerdiagnosing AT naderanmaryam ispolʹzovaniesvertočnyhnejronnyhsetejdlâdiagnostikirakamoločnojželezy AT zaychenkoyuriy ispolʹzovaniesvertočnyhnejronnyhsetejdlâdiagnostikirakamoločnojželezy AT napoliamedeo ispolʹzovaniesvertočnyhnejronnyhsetejdlâdiagnostikirakamoločnojželezy AT naderanmaryam vikoristannâzgortkovihnejronnihmereždlâdíagnostuvannârakumoločnoízalozi AT zaychenkoyuriy vikoristannâzgortkovihnejronnihmereždlâdíagnostuvannârakumoločnoízalozi AT napoliamedeo vikoristannâzgortkovihnejronnihmereždlâdíagnostuvannârakumoločnoízalozi |