Volume 29, Issue 3 (5-2022)
RJMS 2022, 29(3): 240-253 |
Back to browse issues page
Research code: 0
Ethics code: 0
Clinical trials code: 0
Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:
Ghochdashi E, Meshgini S, Makouei S. Breast Cancer Detection from Mammography Images Using Convolutional Neural Network. RJMS 2022; 29 (3) :240-253
URL: http://rjms.iums.ac.ir/article-1-8425-en.html
URL: http://rjms.iums.ac.ir/article-1-8425-en.html
Assistant Professor; Department of Biomedical Engineering, Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran , meshgini@tabrizu.ac.ir
Abstract: (232 Views)
Background & Aims: Mammography is one of the reliable methods for early detection of breast cancer. However, it is difficult for the radiologist to provide an accurate and uniform assessment of the massive mammograms produced in the extensive screening. Therefore, the presence of an intelligent system that is highly accurate in detecting the location of a cancerous mass will be very necessary and necessary. In this regard, this research, by using mammography images and image processing techniques, has been tried to get an accurate diagnosis of the location of breast cancer. For this purpose, first by using some digital image enhancement techniques, an attempt is made to increase the recognition of cancerous tissues, and then by using classification techniques, the precise separation of cancerous parts from healthy parts of the breast is done. In research, various techniques have been proposed to improve the detection of tumors in mammograms and the accuracy of breast cancer classification. The basic problems in breast mammography in identifying and classifying masses and microcalcifications are caused by various factors. One of these complications is due to the awkward and illogical shape of some clusters of calcifications. The boundaries of each of the microcalcifications in the cluster cannot be well defined, and the radiologist may not be able to make an accurate diagnostic decision about the clinical nature of the microcalcifications in an area, but he can usually identify suspicious areas. In the paper, they presented a CAD system for processing mammographic images. They used the compressed breast thickness parameter for feature selection showed the importance of breast compression and changes in breast composition and then applied it to a variety of mammography image processing tasks. Considering that breast thickness is a key parameter in calculations and is not usually recorded; they showed that breast thickness can be estimated from an image and examined its sensitivity on the estimates. Then they discussed how to simulate X-rays in each examination and also simulate the appearance of anatomical structures inside the breast. In the research, tissue characteristics were used to automatically evaluate breast tissue density in digital mammograms. In this approach, the target area is limited to breast tissue only; so that artifacts, background, and head muscles are removed.Breast cancer is the most common cancer among women and the second cause of cancer-related death in women. Mammography is a simple type of imaging and a tool for early detection of non-palpable breast cancers; however, examining and interpreting a large number of mammogram images is a challenging and time-consuming task, and the possibility of human errors is high. One of the most important deep learning methods is convolutional neural networks. In the article, the digital database for screening mammography from the CBIS version was used to improve data validation.
Methods: In this research, three types of architecture were designed in the two-class mode and one type of architecture was designed in the three-class mode. To design the network, the layers were arranged according to Figure 5, which uses an input layer of size 159 x 145 a two-dimensional convolution layer of size 20 x 8, and a maximum integration layer of size 5 x 2, and two fully connected layers. (The maximum integration layer was used because it uses the maximum amount of neuron clusters of the previous layer and also causes faster convergence, and improves generalization and selection of invariant features). The third designed network architecture is shown in such a way that one input layer three 2D convolutional layers three maximum integration layers and two fully connected layers are used, the size of each layer is shown in Table (3). Layering is equal to one. The training time is 6:37 and the accuracy obtained for the validation data is 92.58% and the test data is 86.5%.
Results: The simulation results for 310 data for the second type of two-class architecture, the training time is 6:06, and the accuracy obtained for the validation data is 84.40% and 72.82% for the test data. Also, the simulation results for 1240 The data for the first type of two-class architecture, the training time is 3:44:54, and the accuracy obtained for the validation data is 51.72% and the test data is 51.69%.
Conclusion: After a series of pre-processing, the number of used images was selected as 310. Then two other types of architecture were designed, and by applying the processed data, the accuracy of the architectures for 310 data was 42.39%, 7and 2.82%, respectively. 79.34% was obtained. The accuracy of the architectures for 1240 data was 51.69%, 65.45%, 72.46%. In the three-class mode, 1318 images in the database were used, and due to the lack of the same size, the images were resized. Then the image mask was applied to the images and given to the designed convolutional neural network, and the data was classified into three classes. According to the pre-processing and operations that have been done, the accuracy of the network has increased (72.39%) and the result has improved. The advantage of the method is the increased accuracy of validation and test data.
Methods: In this research, three types of architecture were designed in the two-class mode and one type of architecture was designed in the three-class mode. To design the network, the layers were arranged according to Figure 5, which uses an input layer of size 159 x 145 a two-dimensional convolution layer of size 20 x 8, and a maximum integration layer of size 5 x 2, and two fully connected layers. (The maximum integration layer was used because it uses the maximum amount of neuron clusters of the previous layer and also causes faster convergence, and improves generalization and selection of invariant features). The third designed network architecture is shown in such a way that one input layer three 2D convolutional layers three maximum integration layers and two fully connected layers are used, the size of each layer is shown in Table (3). Layering is equal to one. The training time is 6:37 and the accuracy obtained for the validation data is 92.58% and the test data is 86.5%.
Results: The simulation results for 310 data for the second type of two-class architecture, the training time is 6:06, and the accuracy obtained for the validation data is 84.40% and 72.82% for the test data. Also, the simulation results for 1240 The data for the first type of two-class architecture, the training time is 3:44:54, and the accuracy obtained for the validation data is 51.72% and the test data is 51.69%.
Conclusion: After a series of pre-processing, the number of used images was selected as 310. Then two other types of architecture were designed, and by applying the processed data, the accuracy of the architectures for 310 data was 42.39%, 7and 2.82%, respectively. 79.34% was obtained. The accuracy of the architectures for 1240 data was 51.69%, 65.45%, 72.46%. In the three-class mode, 1318 images in the database were used, and due to the lack of the same size, the images were resized. Then the image mask was applied to the images and given to the designed convolutional neural network, and the data was classified into three classes. According to the pre-processing and operations that have been done, the accuracy of the network has increased (72.39%) and the result has improved. The advantage of the method is the increased accuracy of validation and test data.
Type of Study: Research |
Subject:
Biotechnology
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
