Page 27 - IJEEE-2022-Vol18-ISSUE-1

Atiyah & Ali
A. Dataset
This section discusses publicly accessible datasets used
in this research. For fully automated brain lesion detection
and segmentation, the BraTS2020 dataset is proposed.
BraTS has always been focused on testing advanced methods
for segmenting MRI-based brain tumors. To separate brain
tumors like gliomas, inherently different forms, appearances,
and histologies, BraTS2020 relies on multi-institutional MRI
imagers. BraTS'20 also combines integral radiomic
properties analysis with machine learning methods to
demonstrate the clinical value of this segmentation task by
predicting the patient's general survival and distinguishing
pseudo development from true tumor recurrence. Finally,
BraTS'20 is trying to assess tumor segmentation algorithmic
uncertainty.
Any BraTS multimodal scanning in the nifty form
describing the T1, T1ce, T2, and T2-FLAIR is provided and
has been acquired from a range of clinical protocols and
organizations involved in scanning.
B. U-Net Architecture
In 2015, Ronneberger et al.[11] projected the U-Net
architecture, which was among the first convoluted networks
created exclusively for biological-image interpretation. The
model has a shape similar to the English letter "U." The
encoder is also known as the contracting path, is made up of
the basic convolutional process, whereas the decoder, also
known as the expansive path, is built up of transposed 2D
convolutional layers, as seen Fig. 3.
Each operation in the contracting path contains 2 layers
of convolution, with the number of channels increasing from
1 to 64 as the image depth is enhanced via the convolution
process. The red color arrow pointing down represents the
max-pooling process, which reduces the image size by half
(the size reduction from 572 × 572 ? 568 × 568 is due to
padding difficulties, but padding = “same” is used here).
The image was resized to its original size in the
expansive path. Transposed convolution is a technique for
increasing the size of images by up-sampling them. It
basically pads the original image before doing a convolution
process. The image is upsized from 28×28×1024 to 56×56
×512 after the transposed convolution. The purpose of this
process is to aggregate the information from the previous
layers to obtain a more exact forecast, and two additional
convolution layers are added as well. This technique is
recurrent three times more, as previously. The final stage is
to reform the image to meet our prediction criteria. The
former layer is a convolution layer with one 1×1 filter. The
dense layer, which is particularly prevalent in CNNs for
classification tasks, is not present in the entire network. The
rest of the training for neural networks is the same.
The expansion path in a conventional U-Net is almost
symmetrical to the contracting path. Instead of employing a
standard set of convolution layers, we propose adopting a
novel transfer learning architecture as an encoder in the
contracting path in this work. The decoder module works in
the same way as the original U-Net. To reliably separate the
brain tumor in MR images, we suggest various U-Net
semantic segmentation topologies. The U-Net structure is
| 23
used in these five networks. An encoder (down-sampling
path) extracts image context maps, and a decoder (up-
sampling path) extends the collected mappings for a pixel
segmentation mask to be generated in each network. In
addition, every network features skipping connections that
allow the transmission of information via a precise
segmentation map from the encoding path to the
corresponding levels.
Fig. 3: U-Net Architecture
C. Data Preprocessing
Center cropping is applied to crop the images from the
center. The MRI intensity value relies on the imaging system
and scanner used, standard scaler normalization is employed
to remove the distortion from the image.
Sample ??’s standard score is calculated as follows:
?? = (??-??) (1)
??
Where ?? is the ‘standard deviation’, and ?? is the ‘mean’ of
the training sample. Standard scaler's concept is that it will
transform the data into distribution with a standard deviation
of 1 and a mean of 0. This is done feature-by-feature in the
case of multivariate data.
D. Edge Based Segmentation
To determine the borders of the wanted object within the
image when accessing the image processing, the main
purpose for the edge sensing technique can be defined. The
edge algorithms are done by assessing the sudden increase or
fall in every intensity of the pixels and only visualizing the
sudden changes in the pixels. This pixel difference is crossed
by a suitable convolution mask, and the resulting edges of
the image are described [12].
The approach for the detection of the canny edge is used
in this study. Canny edge detection is one of the most
impactful and frequently used image processing devices for
edge detection. In 1986, John F. Canny developed the canny
edge detection algorithm [13]. This algorithm contains the
basic steps such as reducing noise using gaussian filters,
calculation of gradients along the vertical and horizontal
axes, non-maximal falsified edge removal, double seclusion
thresholds for weak and strong borders, hysteresis edge
tracking.