Page 110 - IJEEE-2022-Vol18-ISSUE-1
P. 110
106 | Majid & Ali
B. Image Augmentation
Large volumes of data are required to create solid and Data splitting with a ratio of 70:30, where training data
is 70%, testing data is 20%, and validation data is 10%.
generalized deep learning models. On the other hand, Data
on medical imaging is limited and difficult to get because of TABLE I
the patient privacy issue; moreover, labeling this data is
expensive. The number of images in the Covid-19 X-rays collection
Fig.4: displays several X-ray images that were not utilized traditional technique.
in the experiment.
Main Dataset Training Validation Testing
set set set
Original dataset 370 176 41
Original dataset +
traditional 2511 717 359
augmentation
2) CGAN Image Augmentation
In this research, the total layers employed in the initial
phase of network generation were four transposed
convolutional layers, followed by three ReLU activation
function layers. For normalizing the values, batch
normalization with three layers and used Tanh activation
Layer for the output. There are four convolutional layers in
the discriminator phase, followed by three leaky ReLU
activation layers. To normalize the values applied two batch
normalization layers and three drops out. Finally, for
the output, used the sigmoid activation Layer. Each
transposed convolutional and convolutional layer seems to
have a 5x5 filter size.
Fig.5: Proposed CGAN and traditional data augmentation Fig.7: layered proposed CGAN generator
are employed in the design stage
1) Traditional Image Augmentation
Traditional Image Augmentation used Random rotation,
scaling, flipping, and shift operations [31][32]. To construct
the Image Data Generator function of the TensorFlow from
the Keras framework [33]. In the traditional augmentation
methodology. COVID-19 X-ray augmented images are
shown in Figure 6.
Fig.6: Classical Augmentation Techniques, which augment Fig.8: layered of proposed CGAN discriminator
the limited COVID-19 chest X-rays images.
CGAN generated sample is assigned the appropriate
label of class (C) and distribution noise (Z). The (G)
generator creates fake images X = G(c,z). Discriminator
(D) produces a probability distribution to class labels and
inputs.
P (S | X), P (C | X) = D(X)
Discriminator layers are initially configured to be non-
trainable. As a result, the discriminator updates the
