Page 29 - IJEEE-2022-Vol18-ISSUE-1

Atiyah & Ali
comparison, elastic transformation, optical distortion to
extend the input picture and provide further information for
the model to be learned. Firstly, before we begin training our
model, we need to establish the learning process. An
optimizer, a loss function, and some metrics such as F1 score
and IoU score have to be specified, optionally. In this study,
the segmentation was based on the U-Net with ResNet50
encoder architecture.
G. Performance Metrics
1) IoU Score: “The intersection of the ground truth with the
prediction segmentation by the place of union between the
ground truth (actual data) and the prediction segmentation
is divided between the ground truth” It is a useful metric to
determine the intersection of two masks or bounding boxes
[14].
?????? = (???????????? ????????h n????????????????????)
(???????????? ????????h ?????????????????????)
2) F1 Score: “The harmonic mean of recall and precision”
is used to get the f1 score[15]. F1 score is also known as
the dice score.
F1 ?????????? = 2 × (?????????????????? × ????????????)
(??????????????????+????????????)
An f1 score can have the highest possible value of 1.0,
which indicates perfect recall and precision, and a lowest
possible value of 0 if either recall or precision is zero.
3) Dice Loss: Prevalent loss function is based on the dice
coefficient for image segmentation tasks. Dice loss is
calculated by subtracting the dice coefficient from 1.
???????? ???????? = 1 - ???????? ??????????????????????
“By multiplying the area of intersection by the total number
of pixels in both images, the dice coefficient is computed”
[16]. The Dice coefficient is calculated as follows:
???????? = 2 × | ??n?? |
|??|+|??|
|AnB| signifies the elements that are common in sets A and
B, and |A| denotes the set A's number of elements. The
same holds true for set B.
When measuring a dice coefficient on predicted
segmentation masks, we may estimate |A n B| as the
element-wise multiplication of the prediction and target
masks and then sum the resultant matrix.
4)Accuracy: Accuracy is the ratio of correctly predicted
observations to the total observations.
???????????????? = (????+????)
(????+????+????+????)
Where TP is True Positive, FP is False Positive, TN is a True
Negative, and FN is False Negative
5) Precision: Precision is the ratio of correctly predicted
positive observations to the total predicted positive
observations.
?????????????????? = ???????? ????????????????
(???????? ????????????????+?????????? ????????????????)
6) Recall (Sensitivity): The recall is the ratio of correctly
predicted positive observations to all positive observations.
| 25
???????????? = ???????? ???????????????? (8)
???????? ????????????????+?????????? ????????????????
7) Specificity: Specificity is the measure of how many true
negatives are predicted out of all actual negative in the
dataset
?????????????????????? = ???????? ???????????????? (9)
???????? ????????????????+?????????? ????????????????
IV. RESULTS & DISCUSSIONS
Segmentation capabilities for the accuracy, dice loss, IoU
score, F1 score, precision, recall, and specificity are analyzed
and implemented in the proposed architectures. The network
is trained in a batch size of 16 for 200 epochs. In this study
comparative analysis of region-based segmentation and
edge-based segmentation using U-Net with ResNet50
encoder, architecture is performed.
The performance metrics of architectures are presented
(2) in Table I. In the edge-based segmentation model, the U-Net
with ResNet-50 encoder architecture achieved the dice loss
score of 0.008768, IoU score of 0.7542, f1 score of 0.9870,
the accuracy of 0.9935, the precision of 0.9852, recall of 0.
9888, and specificity of 0.9951. Figure 6 shows the learning
curve and Fig. 8 shows the prediction of the region-based
segmentation model.
The U-Net with ResNet-50 encoder architecture
achieved the dice loss score of 0.009538, IoU score of
(3) 0.7375, f1 score of 0.9846, the accuracy of 0.9923, precision
of 0.9807, recall of 0.9886, and specificity of 0.9935 in the
region-based segmentation model. Figure 7 shows the
learning curve and Fig. 9 shows the prediction of the region-
based segmentation model. When the results of edge-based
and regional segmentation architecture are compared, edge-
based segmentation is better in all performance metrics.
For training a model, the training parameters are the most
relevant factor. It is therefore essential to use the same
dataset and set all training parameters in the same manner. It
can be used for image segmentation after the network is
trained. The segment images using the trained model take
only a few seconds. On the other hand, it may take hours to
manually segment tumors by clinicians. The image
segmentation procedures suggested helps doctors diagnose a
(4) brain tumor quickly and accurately so that many people can
perhaps save their lives.
Table II illustrates the comparison of the results of the
proposed edge-based segmentation using U-Net with
ResNet50 architecture with some state-of-the-art methods. In
its comparison with the proposed model with the results from
the previous researches, shows that the proposed
(5) segmentation model performed well with the highest
accuracy of 99.35% and a dice score of 98.70%.
This study proposed an automatic method to segment the
brain tumor using a 2D network. So the architecture lacks a
significant degree of semantics and local features between
the pieces. This is the limitation of this study.
(6)
(7)