Page 109 - IJEEE-2022-Vol18-ISSUE-1
P. 109
Majid & Ali | 105
to the generator when the discriminator finds the fake data V. PROPOSED MODEL
[20][21].
The suggested architecture comprises two main
B. Conditional Generative Adversarial Network (CGAN) stages: the first stage is standard data augmentation and the
Conditional GAN (CGAN) [22], a version of the second with CGAN. Later, combined the two techniques and
improve the evaluation. Figure 3 shows the flow chart of the
traditional GAN that employs class labels to condition the proposed model.
class of the generated image from among all the classes in
the training set, allowing for a more realistic image. Suppose
both the G and D are conditioned with additional
information, like class labels. In that case, GAN can be
expanded to a conditional model. Conditioning can be done
by adding the class label y to the Discriminator and
Generator as an extra input layer. The last input, y, and noise
pz(z), are combined with the joint hidden layer in the
generator. The framework for adversarial training networks
provides a lot of freedom in building this private
representation [23].
Fig.2: the CGAN general structure.
The discriminator takes inputs x and y then applies the Fig.3: Proposed Model for data augmentation in three
discriminative. design stages
Equation (2) is the objective function of a two players A. Image Preprocessing
minimax game.
The X-rays images were preprocessed using the same
Min max ?? (D, G) = ????~??????????(??) [log(??(??|??))] + color system, style, and size images were resized to (128 x
128 pixels). Additionally, some of these images were
?? ?? cropped to delete any details not part of the primary X-rays
images like the header and footer. X-ray-generated
????~????(??) [log (1 - ??(??(??|??)))] (2) annotations and arrows were also deleted from the images.
Non-linearity activation Tanh in the CGAN architecture,
? D(x|y) is a discriminator's estimation of the likelihood normalized images to the [-1, 1] range. All of the images in
of actual data example the dataset were treated to remove any scans that were of
? (x) is a reality for a given class (y) poor-quality Figure 3. COVID-19 patients' cleaned X-ray
? The discriminator D(G(z|y)) estimation of the likelihood images. This was done to allow a fair comparison with other
of a fake sample is to be fundamental for class (y) that is standard cases whose data had an Anteroposterior (AP)
given. frontal view of the X-rays images removed from the dataset.
To show the usefulness of the suggested model, we utilized
IV. DATASET’S CHARACTERISTICS just 589 images from the primary datasets in our experiment.
It could only operate with a few images, roughly 10% of the
Used a publicly accessible chest X-ray dataset with three total data. The dataset was updated in 2021[30].
categories: Normal (10,192) images for healthy persons,
COVID-19 (3.616) for patients with a positive case, and
Pneumonia for people with Pneumonia (1,345 images). The
images were varied in size and type [24-29].
