Page 108 - IJEEE-2022-Vol18-ISSUE-1
P. 108
104 | Majid & Ali
overcome this problem, proposed professional Generative (COVID-CGAN). The original dataset included COVID-19,
Adversarial Networks [8],[9]. Especially the conditional- Normal, and pneumonia to CGAN model training. The
Augmentation GAN (CGAN) model [10]. In this research, to limitation: slow it trained timed roughly 16 hours to
create x-rays synthetic COVID-19 images to augmentation complete the training; the proposed model is complex and a
dataset and then utilized in deep detection model, prepared long time to prepare and generate images. Mohd Asyraf et al.
proposed work-based CGAN model to produce synthetic X- [17] creates synthetic data using (DC-GAN). The limitation:
ray including Normal and COVID-19 images. As a result, is that it does not do hyper-parameter tuning. Yifan Jiang et
unsupervised data augmentation was a possible instance al. [18] presents a COVID-19 CT image synthesis approach
when labels were available for a sample of the images in the based on CGAN. A disadvantage is that CT scans were used.
dataset.
III. METHODOLOGY
The following are the most significant contributions
made by this article: A. Overview Generative Adversarial Network (GAN)
• A modified CGAN was built to produce a significant
number of X-ray images to detect COVID-19. This GAN stands for convolution neural network; it is an
comprises the Generator and Discriminator network designs approach for learning deep representations without extensive
also the parameter settings. training data. It was developed by a team of academics
• It was possible to generate Augmented COVID-19 images headed by Ian Good fellow (2014). There are two neural
that included 5589 images that could be utilized to develop network models: generator (G) and discriminator (D). It
the detection models of COVID-19. competed against another in GAN. The noise (Gaussian or
• detection model of COVID-19 using the high-quality standardized distribution) generates samples from the
generated X-rays images was created by the CGAN required distribution [19]. (This is called the generator
modified. model). The discriminator model, which gets samples from
• good detection accuracy was achieved. the generator and training data, is in the second model.
The remainder of our research is organized in the Fig.1: the Generative Adversarial Network's structure
following manner: Section-1 this review provides an
introduction, in section-2 related work review, Methodology Gan has been trained in the manner of a min-max
principle, and some of its modifications in Section-3. algorithm. The loss function is similar to a min-max game
Section-4 Explains the dataset utilized in our research; in with two players, shown by Equation (1).
section-5, there is an overview of the proposed models; in
Section-6, we'll look into how GAN can be used to Min max ?? (D, G) = ????~??????????(??) [log(??(??))] +
synthesize medical images. Section-6 and 7 summarize the
overall review and the result. ?? ??
II. RELATED WORKS ????~????(??) [log (1 - ??(??(??)))] (1)
Receive continued to develop in medical research, • The possibility of instance x being real is predicted by d(x)
especially those interested in Synthetic data generation that
realistic-looking medical images have been proposed in the by a discriminator.
field of healthcare to increase the diversity and quantity of • Ex The average of all actual data.
current training data. Ghorbani et al. [11] propose a synthetic • G(z) the output in the presence of (z) gaussian noise.
data generator based on (GAN) to increase the diversity and • The discriminator D(G(z)) estimate the likelihood of a fake
quantity of skin lesion images. Kohlberger et al. [12]
synthesize pathology images for cancer with natural out-of- sample being actual.
focus characteristics to assess general pathology images for • The predicted value to random input data is given by the
focus quality issues. Han et al. [13] propose that high-
resolution artificial radiographs be created through synthetic generator Ez.
generation. In the space of COVID-19, the authors of the
article Xin Yi et al. [8] show GANs in medical imaging that In Gan, Generator (G) is trained to make an image look
goes beyond image synthesis. And Abdul Waheed et al. [14]
employed the Odena et al. developed (ACGAN) to produce like an original image, while the discriminator discriminates
synthetic data to solve the constraints of traditional
information augmentation. The limitation is Blurry results in between generated and actual data. G produces fake samples
poor image quality. Loey et al. [15] Conditional GAN
(augmentation was utilized to improve multiclass as training samples derived from the latent noise variable (z).
classification accuracy to distinguish COVID-19 from
ordinary. The limitation: limited data was used using only In contrast, fake examples taken from Generator (G) are
345 images. Amal A. Al-Shargabi et al. [16] To exploit the
Covid 19 dataset, used a (CGAN) to build a synthetic image given to the Discriminator (D) to determine the difference
between original and fake data. The error is backpropagated
