Page 111 - IJEEE-2022-Vol18-ISSUE-1

Majid & Ali | 107

generator. This results in a CGAN composite model trained including quicker and earlier identification of any diseases.
to generate fake images for COVID and Non-Covid classes. Through sounds, texts or images, the Deep Learning model
Resized generated images to (128 × 128 × 1) and normalized learns the classification tasks directly. The DL model can
with the range [0, 255] to [1, 1] during the image preparation achieve an accuracy that exceeds the level of human
stage. (The technique of Normalization for changing the performance. The proposed deep learning model detection is
capacity of image pixels.) Its solution is to transform an input a CNN model. In the customized CNN model conation of X-
image into a set of pixels that are more recognizable or rays image as the input layer convolution layers, Relu
natural to the human eye. The optimizer function Adam is activation functions employed after the convolutional layer
employed [34]. Adam is simple, works with sparse gradients, to activate the layers, Max pooling layers, fully connected
takes up minimal memory. As a result, Adam was chosen as layers, and for output layer, sigmoid activation functions are
the optimizer. For CGAN training, the hyperparameters are applied to predict the test image into COVID-19 and Normal,
being used in Table II. which was trained using the optimizer an Adam just ten
epochs, the learning rate is 0.001. The loss function was
TABLE II categorical cross-entropy, all models were trained until
The hyperparameters are being used in the CGAN model. convergence. the network was trained with batch size 32 to
input images with only 6,552,898 parameters,
Hyperparameters values Comparatively, the proposed model utilizes less memory
than ResNet-50, a commonly used deep learning model that
Batch Size 256 employs 23,567,299 parameters for comparison.
However, the DropOut is 0.5, which is essential in avoiding
Learning Rate (LR) 0.0002 overfitting issues of the proposed method.

The Adam optimizer's momentum (Beta) 0.5

number of epochs 20000

Sample interval 5000

The model takes around two hours, forty-seven minutes,
and six-second to train and generate images. Loss functions
optimize the GAN is Binary Cross-Entropy (BCE). in the
design of Table III, all training, validation, and testing data
results are achieved.

TABLE III Fig.9: Model for COVID-19 detection using a proposed
Image numbers of Covid-19 X-rays collection with the CNN.

CGAN technique.

Main Dataset Training Validation Testing
set set set
TABLE IV
The hyperparameters are being used in the Customized

model

Original dataset 370 176 41 Hyperparameters values

optimizer Adam

Original dataset + Learning Rate (LR) 0.001

traditional 2511 717 359 momentum (Beta) 0.5

augmentation number of epochs 10

Batch Size 32

Original dataset+ 1970 563 281 VI. EXPERIMENTAL RESULTS AND DISCUSSION
proposed CGAN
The recommended model was trained on python Google
Original dataset + 4892 1398 698 colab storage than Pro with an increased graphics processing
traditional unit (GPU) and supported GPU learning using the package
and library of TensorFlow machine learning. Colab pro
Augmentation+ Gradient provides paid-tier instances of the NVIDIA M4000
proposed CGAN and NVIDIA P5000 and instances up to V100. The proposed
model has been evaluated in three distinct scenarios: the first
3) Deep Learning Model utilizes the original chest X-rays dataset to augment
traditional augmentation techniques, the second uses
With the aid of medical image classification and the CGAN model, and the third scenario combines the two
detection, Deep Learning models have achieved substantial
advancements in various digital image applications,