Page 274 - 2024-Vol20-Issue2

270 |
Fig. 2. Samples of the dataset: (A) Benign, (B) Malignant,
and (C) Normal.
images was the DICOM format. CT procedure consists of
120 kV, Window sizes of 350 to 1200 HU, slices of 1 mm
of thickness, and window centers ranging from 50 to 600 for
reading with full inspiration, with breath-hold. Before the
analysis process, all photos were undergone de-identification.
Each scan includes a number of slices. The number of these
slices ranges between 80 and 200, and each of them is rep-
resented with a picture of the patient’s chest shown from
different perspectives and various sides. Each of the benign
patients had a different age, gender, geographical location,
educational level, and living condition. Some of these patients
were workers and farmers, while another portion of them were
employees of the Transport and Oil Ministry of Iraq. Mostly
living in the provinces of Babylon, Wasit, Salahuddin, Diyala,
and Baghdad. Fig. 2 shows a sampling of our datasets. The
dataset was divided into three categories: malignant (40), be-
nign (15), and normal (55).
Benign lesions, such as benign tumors or cysts, are non-
cancerous but may resemble cancer on scans. They typically
have well-defined borders and uniform appearances, unlike
malignant tumors, which are often irregular and rapidly chang-
ing. A common benign lung lesion is a hamartoma, character-
ized by its slow growth and well-circumscribed nature [22],
while normal lung imaging shows healthy lung tissue with no
abnormalities. These images are essential for comparison to
identify pathological changes in subsequent scans [23].
C. Preprocessing
In this study, two major preprocessing steps are utilized. First, volutional layer blocks followed by three layers fully linked.
the dataset images came with an original size of 512x512 pix-
els; the size of these images was reduced to 224x224, which
is the default input size of the VGG16 network. Second,
Qadir, Abdalla & Abd
all of the pixels were normalized to the range of 0 and 1 in
order to reduce the computational complexity of the image
recognition process. In our study, we addressed the challenge
of class imbalance in our dataset through the application of
Stratified 5-fold Cross-Validation. This technique is particu-
larly beneficial for imbalanced datasets as it ensures that each
fold maintains a consistent ratio of the different class labels,
closely mirroring the original dataset’s class distribution. By
employing this method, we aimed to achieve a more reliable
and unbiased evaluation of our model’s performance. Each
of the five folds was used once as a validation set, while the
remaining four folds served as the training set. This approach
not only allowed for a thorough assessment of the model
across different subsets of the data but also contributed to miti-
gating the potential bias that could arise from the uneven class
distribution. The consistent representation of classes in each
fold helped in validating the robustness and generalizability of
our model, ensuring that the validation metrics were reflective
of its performance across the diverse class spectrum [24].
D. Feature Extraction using VGG-16
Initially, the first VGG network design was proposed in a
study by Simonyan and Zisserman [25]. Two variations of the
model were created, which are a 16-layer (VGG-16) as well
as a 19-layer (VGG-19) network, with the aim of submitting
these models and participating in the ImageNet challenge in
2014, where the competition was won by the team of Visual
Geometry Group (VGG), achieving a second place in the lo-
calization track while taking first place in the classification
track. The architecture of VGG16 is composed of five con-
Convolutional layers make sure that each of the activation
maps stays at the same spatial dimensions compared to the
layer, which is below it, through the use of 3x3 kernels and a
padding and stride of 1.
After each convolution, a rectified linear unit (ReLU) activa-
tion and max pooling operation are immediately performed
to reduce the spatial dimension. In the activation map, each
spatial dimension from the previous layer is cut in half when
using 22 kernels through the max pooling layers while using
two strides and without padding. The final layer is the softmax
layer, which consists of 1,000 completely linked layers and
is then employed after two fully connected layers containing
4,096 units with a ReLu activation function.
At the same time, one of the drawbacks of the VGG6 model is
that it is expensive to analyze, and it requires a lot of process-
ing power in terms of memories and parameters. The model
contains almost 138 million parameters. In the proposed
model, an XGBoost classifier replaces the fully connected
layers, in which approximately 123 million parameters reside.
These layers are considered the bulk of these parameters; this