Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози
In this paper, the breast cancer detection problem using convolutional neural networks (CNN) is considered. The review of known works in this field is presented and analysed. Most of them rely only on feature extraction after the convolutions and use the precision of classification of malignant tumo...
Збережено в:
| Дата: | 2022 |
|---|---|
| Автори: | , , |
| Формат: | Стаття |
| Мова: | Англійська |
| Опубліковано: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2022
|
| Теми: | |
| Онлайн доступ: | https://journal.iasa.kpi.ua/article/view/265629 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | System research and information technologies |
| Завантажити файл: |  |
Репозитарії
System research and information technologies| _version_ | 1867334428497281024 |
|---|---|
| author | Zaychenko, Yuriy Naderan, Maryam Hamidov, Galib |
| author_facet | Zaychenko, Yuriy Naderan, Maryam Hamidov, Galib |
| author_institution_txt_mv | [
{
"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": "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": "Galib Hamidov",
"institution": "“Azershig” company, Baku"
}
] |
| author_sort | Zaychenko, Yuriy |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2022-10-17T22:12:39Z |
| description | In this paper, the breast cancer detection problem using convolutional neural networks (CNN) is considered. The review of known works in this field is presented and analysed. Most of them rely only on feature extraction after the convolutions and use the precision of classification of malignant tumors as the main criterion. However, because of the huge number of parameters in the models, the time of computation is very large. A new structure of CNN is developed — a hybrid convolutional network consisting of convolutional encoder for features extraction and reduction of the complexity of the model and CNN for classification of tumors. As a result, it prevented overfitting the model and reduced training time. Further, while evaluating the performance of the convolutional model, it was suggested to consider recall and precision criteria instead of only accuracy like other works. The investigations of the suggested hybrid CNN were performed and compared with known results. After experiments, it was established the proposed hybrid convolutional network has shown high performance with sensitivity, precision, and accuracy of 93,50%, 91,60%, and 93%, respectively, and requires much less training time in the problem of breast cancer detection as compared with known works. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2022.2.06 |
| first_indexed | 2025-07-17T10:27:58Z |
| format | Article |
| fulltext |
Yu. Zaychenko, M. Naderan, G. Hamidov, 2022
Системні дослідження та інформаційні технології, 2022, № 2 85
TIДC
ТЕОРЕТИЧНІ ТА ПРИКЛАДНІ ПРОБЛЕМИ
ІНТЕЛЕКТУАЛЬНИХ СИСТЕМ ПІДТРИМАННЯ
ПРИЙНЯТТЯ РІШЕНЬ
UDC 519.925.51
DOI: 10.20535/SRIT.2308-8893.2022.2.06
HYBRID CONVOLUTION NETWORK FOR MEDICAL IMAGES
PROCESSING AND BREAST CANCER DETECTION
Yu. ZAYCHENKO, M. NADERAN, G. HAMIDOV
Abstract. In this paper, the breast cancer detection problem using convolutional
neural networks (CNN) is considered. The review of known works in this field is
presented and analysed. Most of them rely only on feature extraction after the con-
volutions and use the precision of classification of malignant tumors as the main cri-
terion. However, because of the huge number of parameters in the models, the time
of computation is very large. A new structure of CNN is developed — a hybrid con-
volutional network consisting of convolutional encoder for features extraction and
reduction of the complexity of the model and CNN for classification of tumors. As
a result, it prevented overfitting the model and reduced training time. Further, while
evaluating the performance of the convolutional model, it was suggested to consider
recall and precision criteria instead of only accuracy like other works. The investiga-
tions of the suggested hybrid CNN were performed and compared with known re-
sults. After experiments, it was established the proposed hybrid convolutional net-
work has shown high performance with sensitivity, precision, and accuracy of
93,50%, 91,60%, and 93%, respectively, and requires much less training time in the
problem of breast cancer detection as compared with known works.
Keywords: breast cancer detection, hybrid convolutional network, encoder, classifi-
cation sensitivity, dimensionality reduction.
INTRODUCTION
Breast cancer is a very common cancer among women between the ages of 35 and
55 [1]. Diagnosing breast cancer is frequently discussed as a classification prob-
lem within neural networks. Detecting and diagnosing breast cancer in early
stages is critical in saving women’s lives. Detecting this cancer in its early stages
can help prevent the spread of cancer to other organs/tissues allowing doctors to
help the patient before it is too late. Early detection requires methods that are sys-
tematic and dependable, allowing healthcare professionals to accurately distin-
guish between benign and malignant tumors [2]. For these reasons, the exact de-
tection and classification of breast tumors is extremely important for public health
and to the lives of cancer patients [3].
There are four types of breast cancer: in situ, invasive ductal carcinoma, in-
flammatory breast cancer, and metastatic cancer [4]. Breast cancer detection is
important in developing countries, where the number of patients is dramatically
Yu. Zaychenko, M. Naderan, G. Hamidov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 86
higher. Moreover, detecting breast cancer is a challenging and time-consuming
task requiring doctors to manually label scans. Although, there are supervised and
unsupervised machine learning algorithms that assist doctors. According to the
World Health Organization (WHO) [5], a mammography scan is more efficient
and cost-effective for breast cancer detection. Although it is more expensive than
other medical images, the quality of the image is superior to other medical scans
and therefore mammography scans, from the BreakHis dataset [6] have been used
in this work.
The main goal of this work is the development and investigation of hybrid
convolutional network to increase the sensitivity and to reduce the complexity of
the model for breast cancer detection. A convolutional autoencoder was proposed
to extremely decrease the computation time.
REVIEW OF PREVIOUS WORKS
There are a lot of studies that consider breast cancer detection using CNNs. How-
ever, most of them rely on the accuracy in their experiments, but accuracy in any
cancer detection is not the only valid factor that should be considered [7]. In these
tasks, the sensitivity of models should be considered to understand how many
times the model misclassified cancer. Authors in [8] proposed a state-of-the-art
convolutional neural network (DenseNet) for breast cancer detection using Breast
Cancer Histology images (BACH) with an accuracy of 85,6%. The misclassifica-
tion rate for cancer class was 14,4% on average. In their work, the sensitivity (on
average for 4 classes) of ResNet 50 was compared with their proposed CNN at
76% and 79% respectively.
Compared with pre-existing CNN models (VGG-16, VGG19, Xception,
Resnet, Inception) with 80% accuracy in multiclass classification, authors in [9]
proposed a model where the accuracy was 83,97% on average for two classes
(Benign and Malignant). The proposed model was a combination of Inception and
Resnet using the BreakHis data set, which contains 7909 mammography scans
with four magnification factors ( 40X, 100X, 200X and 400X).
In [10], it was stated that because of the architecture of DenseNet, in which
all layers are fully connected to every previous layer, and with a short connection
between those layers near the input and output, the model could be trained more
efficiently and accurately.
In [11] a DenseNet network authors proposed model which achieved high
processing performances with 95,4% of accuracy. The Authors claim they first
used weights from Imagenet and fine-tuned the model to train DenseNet. All con-
volutional parts of the network were frozen but the fully connected layer was
trainable. Authors in [12] used an atrous DenseNet that achieves multi-scale fea-
ture extraction by integrating the atrous convolutions to the dense block. The au-
thors in [12] compared two datasets, BACH and CCG, in which the average class
accuracy for the proposed model was 82,50% and 87,05% respectively for each
dataset.
A new model of convolutional neural network was proposed in [13], where
the authors used 400 images with 40x magnification for training data and 200 for
validation data. In [13] three different ConvNet architectures were evaluated: 1) a
3-layer ConvNet architecture, 2) a 4-layer ConvNet architecture, and 3) a deeper
Hybrid convolution network for medical images processing and breast cancer detection
Системні дослідження та інформаційні технології, 2022, № 2 87
6-layer ConvNet architecture. The 3-layer ConvNet included one convolution,
one pooling and one fully connected layer. The 4-layer had two convolutional and
two pooling layers and the last layer was fully connected. The 6-layer ConvNet
architecture comprises four convolutional and pooling layers with 16 units, a fully
connected layer. According to the results in [13], deep architectures shows better
result with 1,06% accuracy.
Authors in [14] proposed semi-supervised learning (SSL) using convolu-
tional neural networks. The accuracy of the developed model was 82,43% and the
area under the curve (AUC) observed in their study was 88,18%. There were 1874
pairs of mammogram images used during the experiments. Moreover, the authors
developed three data weighing equations using exponential function, Gaussian
function, and Laplacian function. Based on results [14], comparing two other
weighting equation, the exponential function has shown better results with
82,43%, 81,00% and 72,26% for labelling accuracy, sensitivity and specificity
respectively.
In [16] authors applied Principal component analysis (PCA) for Hybrid
Fuzzy CNN Network. The idea of using PCA was to reduce the number of extracted
features. In their work, the authors proposed a model where CNN VGG 16 was
used for feature extraction and FNN NEFClass was used for image classification.
DATASET
The open source BreakHis dataset was used during the experiment. The dataset
includes two classes benign and malignant tumors. The dataset is also separated
into four magnification zooms 40X, 100X, 200X and 400X; 5000 images were
used for training and 350 images were used for testing. Fig. 1 illustrates some in-
put images that were used for training the model. Fig. 1, a–d belong to the benign
category and Fig. 1, e–h belong to the malignant category.
ARCHITECTURE AND TRAINING OF CONVOLUTIONAL AUTOENCODER
The aim of the autoencoder is to learn a compressed distributed representation for
the given data typically for the purpose of dimensionality reduction. On the other
Fig. 1. Sample of input images: a — adenosis; b — fibroadenoma; c — phyl-
lodes_tumor; d — Tubular_adenoma; e — Ductal_carcioma; f —Lobular_ carcioma;
g — Muconous carcinoma; h — Papillary carcioma
a b c d
e f g h
Yu. Zaychenko, M. Naderan, G. Hamidov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 88
hand, there is a principal component analysis (PCA) for the same task (reduction
dimensionality). However, there are some advantages [17] of using autoencoder
like: 1) autoencoder can represent both linear and non-linear transformations in
encoding but PCA can perform only linear transformations; 2) it could be more
efficient in terms of model parameters to learn several layers with an autoencoder
rather than one massive transformation with PCA; 3) it gives a representation as
the output of each layer and having multiple representation of different dimen-
sions is more practical.
One of the reasons a convolutional autoencoder was used during experi-
ments is because it is very challenging to find significantly sized datasets with
labels and autoencoder is an unsupervised model that does not require the dataset
to be labelled. Another advantage of the autoencoder is that it makes the model
smaller. Respectively, the model would have less parameters and as a result, the
time of computation and training will drastically decrease. For example, in Dense
Net there are a total of 58,420,802 parameters and 7,037,504 of them are not
trainable. However, in the proposed convolutional autoencoder there are
2,940,865 parameters and only 3,840 of them are non-trainable.
Fig. 2 illustrates the architecture of an autoencoder. In the autoencoder there
are layers between the input and output and the sizes of these layers are smaller
than the input layer. For example, the input vector has a dimensionality of N
which means that the output will also have a dimensionality of N. The input goes
through a layer of size P, where the value of P is less than N. The autoencoder
receives unlabelled input which is then encoded to reconstruct input. The impor-
tant part of autoencoder is the Bottleneck approach for representation learning.
In the current work, several architectures of convolutional autoencoder were
used during the experiment. The convolutional autoencoder was modified with 18
encoding layers and 14 decoding layers. There were eight convolutional and two
max pooling layers in encoder. In decoder there were six convolutional and two
upsampling layers. Batch normalization was used between each convolutional
layer. The proposed convolutional autoencoder was trained in a way, that the
model would extract informative features (Codes) during the encoding process,
and the decoder could then reconstruct the original input image of the encoder.
The model could recreate the original image, even though some noises were ap-
plied to the scans. A comparison of the input images and reconstructed images is
shown in Fig. 3. After creating a successful autoencoder-model, the output of the
encoder will be used with a fully connected layer to create a full model (Convolu-
tional Autoencoder).
Latent Space
Representation
Input Image Reconstructed Image
Fig. 2. The architecture of an autoencoder
Hybrid convolution network for medical images processing and breast cancer detection
Системні дослідження та інформаційні технології, 2022, № 2 89
The accuracy of the recreated test data for convolutional autoencoder was
79,38%.
EXPERIMENTAL INVESTIGATIONS AND ANALYSIS
For training an autoencoder there are four parameters that it is needed to be set.
The first one is code size. The code size represents the number of nodes in the
middle layer and smaller size results in more compression. The second parameter
is the number of layers and the autoencoder could be as deep as we want it to be.
Another parameter is the loss function. The last parameter is the number of nodes
per layer. The number of nodes per layer decreases with each subsequent layer of
the encoder and increases back in the decoder. Also, the decoder is symmetric to
the encoder in terms of layer structure.
The Adam optimizer with learning rate 0,001 was used for training Dense-
Net whereas, in convolutional autoencoder the RMSprop ( 001,0rl ) has shown
better results.
All scans were pre-processed, before being used to train the model by resiz-
ing, normalizing and dimensionality reduction methods. In this paper, all experi-
ments were developed using Jupyter Labs, Tensorflow 2 and Python 3. The pro-
grams were implemented on a virtual machine with an NVIDIA Tesla GPU and
eight Intel CPUs.
Fig. 4, a and b illustrate how the loss for training and validation data was
changed. Multiple tests were done using different numbers of epochs. It was
found that 250 epochs provided better results compared to larger number of
a
b
Fig. 3. On scans (a) noises were added, and scans (b) are reconstructed to original
test data
Yu. Zaychenko, M. Naderan, G. Hamidov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 90
epochs like 500. The more epochs, the more chance the model will have overfit-
ting. According to Fig. 6, it is better to use learning rate 0,001 for the current
model. However, this parameter could be different for other models.
In previous work [17], a modified Inception V3 was proposed for breast can-
cer detection. In this work a hybrid convolutional network was proposed using a
fine-tuned DenseNet121 and modified convolutional autoencoder. In the proposed
hybrid convolutional network, the convolutional autoencoder was used as a fea-
ture extraction and the DenseNet was used as a classifier. Fig. 5 demonstrates the
architecture of the proposed model.
Table 1 shows the result of the proposed model for each class.
T a b l e 1 . The results of breast cancer recognition using a convolutional
autoencoder
Class Precision, % Recall, % F1-Score, % Support
Class 0 90 95 93 747
Class 1 95 90 92 1626
Weighted ayg 93,2 93,5 93,3 2373
The experiment proves that it is not necessary to have large data sets to train
a convolutional autoencoder from scratch. Comparatively, training the DenseNet
Encoded
vector
Convolutiional
layer 1
Pooling 1
Convolutiional
layer n
Pooling n
Softmax Layer
(Classifier)
p(y|x)
Cancer
Output
No-
Cancer
Fig. 5. Modified architecture of hybrid convolutional network
Fig. 4. Loss comparison for training and validation data with learning rate 0,01 and 0,001
on (a) and (b) respectively
2
1
1 –
2 –
0,45
0,40
0,35
0,30
0,25
0,20
0,15
a
1 –
2 –
2
1
0,25
0,20
0,15
0,10
0,05
0,00
b
Hybrid convolution network for medical images processing and breast cancer detection
Системні дослідження та інформаційні технології, 2022, № 2 91
and Inception-v3 convolutional networks from scratch, require a large number of
input images. Thus, the convolutional autoencoder has a simplified model, and
training time is significantly reduced compared to DenseNet or Inception-v3. Fig.
6 illustrates the performance of the model with different number of input data as
training data.
Also, the appearance and image quality of the input data significantly affect
the performance of the model. BreakHis and Breast histology datasets were used
for comparison. In the table 2 shows the performance of the model with different
input data.
T a b l e 2 . Comparison of the quality of the model of the convolutional autoen-
coder for different datasets
Factors/Datasets BreakHis, % Breast histology, %
Accuracy 93 90,5
Precision 93,2 91,6
Recall 93,5 92,40
F1-Score 93,3 92
Table 3 shows three different deep convolutional networks that were used
for the current task. According to table 3, the hybrid convolutional network has
shown better results as compared to other methods, the training time was lower.
There are plenty of studies for breast cancer detection using deep learning, and
most of them rely only on accuracy. However, sensitivity (recall) is the most im-
portant factor that should be kept in mind while training deep neural networks.
With recall it is possible to assess whether the network predicts cancer as a can-
cer, so in the current work both recall and precision of the model were considered.
T a b l e 3 . Result of comparison different models for detecting breast
Factors
CNN models
Precision, % Recall, % F1 score, %Training Time
Modified Inception V3 [17] 66,66 85,70 74,99 27h
DenseNet 121 75,73 86,1 80,84 24h
Hybrid convolutional neural network 91,60 93,50 92,5 13h
Fig. 6. Accuracy of proposed model for different size of data sets
, ,
, , , ,
, ,
, ,
,
Accuracy of Convolutional Autonecoder
using different number of training dateset
Yu. Zaychenko, M. Naderan, G. Hamidov
ISSN 1681–6048 System Research & Information Technologies, 2022, № 2 92
Based on the data from table 3 and table 1, it can be concluded that the sensi-
tivity of the model (recall) when using a convolutional autoencoder gives a better
result compared to Inception-v3. It should also be noted that only 5% of class 0
(cancer), was misclassified.
CONCLUSION
1. In this paper, a hybrid convolutional neural network was developed, and
investigated in the problem of breast cancer detection. In the proposed hybrid
convolutional network the convolutional autoencoder was used as a feature ex-
traction while CNN DenseNet was used as a classifier.
2. In the experiments it was determined that sensitivity, precision and accu-
racy of the proposed model were 93,50%, 91,60% and 93% respectively. The
comparison with known CNN was performed which has shown the proposed hy-
brid CNN has higher sensitivity (recall) than known CNN models.
3. Besides the hybrid CNN has fewer parameters compared to DenseNet, as
a result, the model is less complex and prevents overfitting. Moreover, the used
autoencoder is an unsupervised model and does not require labeled data.
4. In addition, during the experiments it was established that hybrid CNN
requires less training time as compared with known CNN models.
REFERENCES
1. K. Polat et al., “A new classification method for breast cancer diagnosis: feature se-
lection artificial immune recognition system (FS-AIRS)”, Advances in Natural Com-
putation, pp. 830–838, 2005.
2. M.R. Senapati et al., “Local linear wavelet neural network for breast cancer recogni-
tion”, Neural Computing and Applications, 22 (1), pp. 125–131, 2013.
3. A. Marcano-Cedeno, J. Quintanilla Dominguez, and D. Andina, “Breast cancer clas-
sification applying the artificial metaplasticity algorithm”, Neurocomputing, 74 (8),
pp. 1243–1250, 2011.
4. Diagnostic Mammogram. Available: https://www.nationalbreastcancer.org/ diagnos-
ticmammogram.
5. Breast cancer: prevention andcontrol. Available: https://www.who.int/news-room/
fact-sheets/ detail/breast-cancer.
6. Breast Cancer Histopathological Database (BreakHis). Available: https://web.inf.
ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/.
7. Inverse Relationship Between Precisionand Recall. Available: https://datascience.
stackexchange.com/questions/49117/inverserelationship-between-precision-and-recall
8. G. Arestaa et al., “BACH: grand challenge on breast cancer histology images”,
Medical Image Analysis, ELSEVIER 2019.
9. BreakHist-Dataset-Image-Classification. Available: https://github.com/Anki0909/
BreakHist-Dataset-Image-Classification
10. Gao Huang, Zhuang Liu, and Kilian Q. Weinberger, “Densely Connected Convolu-
tional Networks”, arXiv:1608.06993v3, 2018.
11. Majid Nawaz, Adel A. Sewissy, and Taysir Hassan A. Soliman, “MultiClass Breast
Cancer Classification using Deep Learning Convolutional Neural Network”, Interna-
tional Journal of Advanced Computer Science and Applications, vol. 9, no. 6, 2018.
12. Yuexiang Li, Xinpeng Xie, Linlin Shen, and Shaoxiong Liu, “Reversed Active
Learning based Atrous DenseNet for Pathological Image Classification”,
arXiv:1807.02420, 2018.
Hybrid convolution network for medical images processing and breast cancer detection
Системні дослідження та інформаційні технології, 2022, № 2 93
13. A. Cruz-Roa et al., “Accurate and reproducible invasive breast cancer detection in
whole-slide images: A Deep Learning approach for quantifying tumor extent”, Sci.
Rep., 7, 46450, 2017. doi: 10.1038/srep46450.
14. Wenqing Sun, Tzu Liang (Bill) Tseng, Jianying Zhang, and Wei Qian (2016). “En-
hancing deep convolutional neural network scheme for breast cancer diagnosis with
unlabeled data”, Computerized Medical Imaging and Graphics. doi:
10.1016/j.compmedimag.2016.07.004
15. Comparison PCA and Autoencoder. Available: https://towardsdatascience.com/ di-
mensionality-reduction-pca-versus-autoencoders-338fcaf3297d
16. Y. Zaychenko and G. Hamidov, “Hybrid Fuzzy CNN Network in the Problem of
Medical Images Classification and Diagnostics”, in Advances in Natural Computa-
tion, Fuzzy Systems and Knowledge Discovery, ICNC-FSKD 2019, Advances in In-
telligent Systems and Computing, vol. 1074. Springer, Cham, 2020.
17. M. Naderan, Yu. Zaychenko, and A. Napoli, “Using convolutional neural networks
for breast cancer diagnosing”, System Research and Information Technologies, no. 4,
pp. 85–93, 2019.
18. M. Naderan, “Review methods for breast cancer detection using artificial intelli-
gence and deep learning methods”, System Research & Information Technologies,
no. 1, pp. 98–101, 2021.
Received 09.11.2021
INFORMATION ON THE ARTICLE
Yuriy P. Zaychenko, ORCID: 0000-0001-9662-3269, Institute for Applied System
Analysis of the National Technical University of Ukraine “Igor Sikorsky Kyiv
Polytechnic Institute”, Ukraine, e-mail:zaychenkoyuri@ukr.net
Maryam Naderan, Institute for Applied System Analysis of the National Technical
University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail:
ma.nederan@gmail.com
Galib Hamidov, “Azerishiq”, Azerbaijan, e-mail: galib.hamidov@gmail.com
ГІБРИДНА ЗГОРТКОВА МЕРЕЖА ДЛЯ ОБРОБЛЕННЯ МЕДИЧНИХ
ЗОБРАЖЕНЬ ТА ВИЯВЛЕННЯ РАКУ МОЛОЧНОЇ ЗАЛОЗИ / Ю.П. Зайченко,
М. Надеран, Г. Гамідов
Анотація. Розглянуто проблему виявлення раку молочної залози з викори-
станням згорткових нейронних мереж (ЗНМ). Наведено огляд та аналіз праць з
цієї галузі. Зазначається, що більшість з них засновано на вилученні ознак у
результаті згортки з використанням як основного критерію точність
класифікації пухлин. Унаслідок великого обсягу параметрів, що
оптимізуються, час навчання дуже тривалий. Розроблено нову структуру ЗНМ
— гібридну мережу, що складається з енкодера для отримання первинних
ознак і скорочення розмірності моделі та декількох шарів згортки для класифі-
кації пухлин. Це дало змогу запобігти перенавченню мережі та скоротити час
навчання. Для оцінювання якості класифікації запропоновано використовувати
критерій чутливості (до злоякісних пухлин) разом із критерієм точності на
відміну від відомих праць. Це дозволило скоротити відсоток пропуску
злоякісних пухлин. Проведено експериментальні дослідження розробленої
гібридної згорткової мережі та порівняно з іншими працями. Установлено, що
гібридна ЗНМ має високі показники якості класифікації, а також чутливість до
ракових пухлин і точність класифікації 93,50%, 91,60% відповідно і потребує
значно менше часу на навчання класифікації пухлин молочної залози порівня-
но з відомими працями.
Ключові слова: виявлення раку молочної залози, гібридна згорткова мережа,
кодер, чутливість класифікації, зменшення розмірності.
|
| id | journaliasakpiua-article-265629 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:27:58Z |
| publishDate | 2022 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/84/9810dd30b33ef0d74ba938190e1b2b84.pdf |
| spelling | journaliasakpiua-article-2656292022-10-17T22:12:39Z Hybrid convolution network for medical images processing and breast cancer detection Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози Zaychenko, Yuriy Naderan, Maryam Hamidov, Galib виявлення раку молочної залози гібридна згорткова мережа кодер чутливість класифікації зменшення розмірності breast cancer detection hybrid convolutional network encoder classification sensitivity dimensionality reduction In this paper, the breast cancer detection problem using convolutional neural networks (CNN) is considered. The review of known works in this field is presented and analysed. Most of them rely only on feature extraction after the convolutions and use the precision of classification of malignant tumors as the main criterion. However, because of the huge number of parameters in the models, the time of computation is very large. A new structure of CNN is developed — a hybrid convolutional network consisting of convolutional encoder for features extraction and reduction of the complexity of the model and CNN for classification of tumors. As a result, it prevented overfitting the model and reduced training time. Further, while evaluating the performance of the convolutional model, it was suggested to consider recall and precision criteria instead of only accuracy like other works. The investigations of the suggested hybrid CNN were performed and compared with known results. After experiments, it was established the proposed hybrid convolutional network has shown high performance with sensitivity, precision, and accuracy of 93,50%, 91,60%, and 93%, respectively, and requires much less training time in the problem of breast cancer detection as compared with known works. Розглянуто проблему виявлення раку молочної залози з використанням згорткових нейронних мереж (ЗНМ). Наведено огляд та аналіз праць з цієї галузі. Зазначається, що більшість з них засновано на вилученні ознак у результаті згортки з використанням як основного критерію точність класифікації пухлин. Унаслідок великого обсягу параметрів, що оптимізуються, час навчання дуже тривалий. Розроблено нову структуру ЗНМ — гібридну мережу, що складається з енкодера для отримання первинних ознак і скорочення розмірності моделі та декількох шарів згортки для класифікації пухлин. Це дало змогу запобігти перенавченню мережі та скоротити час навчання. Для оцінювання якості класифікації запропоновано використовувати критерій чутливості (до злоякісних пухлин) разом із критерієм точності на відміну від відомих праць. Це дозволило скоротити відсоток пропуску злоякісних пухлин. Проведено експериментальні дослідження розробленої гібридної згорткової мережі та порівняно з іншими працями. Установлено, що гібридна ЗНМ має високі показники якості класифікації, а також чутливість до ракових пухлин і точність класифікації 93,50%, 91,60% відповідно і потребує значно менше часу на навчання класифікації пухлин молочної залози порівняно з відомими працями. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2022-08-30 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/265629 10.20535/SRIT.2308-8893.2022.2.06 System research and information technologies; No. 2 (2022); 85-93 Системные исследования и информационные технологии; № 2 (2022); 85-93 Системні дослідження та інформаційні технології; № 2 (2022); 85-93 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/265629/261672 |
| spellingShingle | виявлення раку молочної залози гібридна згорткова мережа кодер чутливість класифікації зменшення розмірності Zaychenko, Yuriy Naderan, Maryam Hamidov, Galib Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| title | Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| title_alt | Hybrid convolution network for medical images processing and breast cancer detection |
| title_full | Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| title_fullStr | Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| title_full_unstemmed | Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| title_short | Гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| title_sort | гібридна згорткова мережа для оброблення медичних зображень та виявлення раку молочної залози |
| topic | виявлення раку молочної залози гібридна згорткова мережа кодер чутливість класифікації зменшення розмірності |
| topic_facet | виявлення раку молочної залози гібридна згорткова мережа кодер чутливість класифікації зменшення розмірності breast cancer detection hybrid convolutional network encoder classification sensitivity dimensionality reduction |
| url | https://journal.iasa.kpi.ua/article/view/265629 |
| work_keys_str_mv | AT zaychenkoyuriy hybridconvolutionnetworkformedicalimagesprocessingandbreastcancerdetection AT naderanmaryam hybridconvolutionnetworkformedicalimagesprocessingandbreastcancerdetection AT hamidovgalib hybridconvolutionnetworkformedicalimagesprocessingandbreastcancerdetection AT zaychenkoyuriy gíbridnazgortkovamerežadlâobroblennâmedičnihzobraženʹtaviâvlennârakumoločnoízalozi AT naderanmaryam gíbridnazgortkovamerežadlâobroblennâmedičnihzobraženʹtaviâvlennârakumoločnoízalozi AT hamidovgalib gíbridnazgortkovamerežadlâobroblennâmedičnihzobraženʹtaviâvlennârakumoločnoízalozi |