![](bg8.png)
layer (in pink color) instead of max pooling in the
decoding (up-sampling) part. The number of filter
channels and the image size are given at the bottom of
each layer. The size of the input layer (in white color)
is 256 × 256 × 3 and the size of the output layer is
256 × 256 × 1 which is a convolution layer with Sig-
moid activation function.
4 EXPERIMENTAL RESULTS
In order to evaluate the performance of the proposed
method, various experiments have been performed on
a challenging MRI dataset. We have used the follow-
ing libraries for the implementation: OpenCV, Pillow,
NumPy, Matplotlib, and Tensorbord for visualization.
The operating system used was Ubuntu 18.04 on a
computer with 5 cores and with an Nvidia GeForce
GTX 960M graphics processor equipped of 9 GB of
RAM. This section includes qualitative and quantita-
tive assessment of the proposed method, while com-
prehensively assesses each module of the method.
4.1 Qualitative Evaluation
To illustrate the performance of the segmentation
model, overlays of FLAIR MRI images with the out-
lines of tumor masks using manual and model seg-
mentations for the test datasets are shown in Figure
8. Each panel, in Figure 8, is showing the highlighted
tumor in red, while the overlay image with tumor out-
lines (green – manual segmentation and red – model
segmentation). Overall, the visualization of the re-
sults allowed us to see that U-Net and SegNet make
complementary errors. Representative examples of
automatic segmentation results obtained using Seg-
Net architecture with best and worst scores are shown
in Figure 9 and Figure 10, respectively. The results re-
flect the anomalous detection of regions, while noise
and random speckles (red dots) indicate that SegNet
tends to miss finer details. It is clear that the proposed
model lacks precision, although it considers very deep
with 10 encoder layers. However, the proposed model
has succeeded in detecting the glioma region even in
the worst case, but it lacks precision. This leads us
to think that it lacks regularization to fit the proposed
problem. This will be discussed in the next part. In
fact, while visualizing the results, we have observed
that the proposed U-SegNet architecture captures fine
details and solves the random noise problem seen in
SegNet as illustrated in Figure 11 and Figure 12. It
is obvious that adding skip connections to the upper
layers helps to improve performance.
Consequently, SegNet tends to miss the finer de-
tails and in some cases suffers from random noise. On
the other hand, U-Net, thanks to jump connections, is
able to capture fine details; i.e. borders; more accu-
rately than SegNet. However, as shown in the same
figure (Figure 8), U-Net makes some errors in the de-
tection of tumors. We suspect this is due to confu-
sion created by deconvolutional layers and skipped
connections at lower levels. Moreover, compared to
U-Net, U-SegNet has fewer parameters than U-Net
allowing our network to train better. This solves the
accuracy problem. Although SegNet tends not to have
access to finer details, the proposed model is able to
capture these finer details by integrating the single
hop connection into the U-SegNet architecture.
4.2 Quantitative Evaluation
To compare the quantitative performances of the dif-
ferent models, we have evaluated the performance of
these segmentations through the Dice similarity (DC)
coefficient. It is among the most widely used met-
rics for brain tumor and structure segmentation appli-
cations. The Dice coefficient (1) was used to evalu-
ate the similarity of the predicted tumor masks by the
segmentation model with the tumor masks obtained
by manual segmentation (GT).
DC =
2 × T P
2 × T P + FN + FP,
(1)
where, TP, FP, and FN represent respectively the
True Positive, False Positive, and False Negative of
the class for which the result is calculated.
Table 1 shows the training time, best Dice coef-
ficient, mean Dice coefficient, and median Dice co-
efficient of each model per 100 epochs. As shown
in Table 1, SegNet performs faster than other models
since SegNet uses only max-pooling indices to over-
sample low-level features. It is obvious that adding
skip connections to the upper layers helps to improve
the performance. Thus, U-SegNet gave an average
Dice value of 83% and a median Dice coefficient
of 86%. Network training required 8 GB of mem-
ory while the total training time was approximately
5 hours and 58 minutes. In Figure 13, we present
the loss and Dice convergence results of the valida-
tion dataset for each of these models. Both U-Net and
U-SegNet models seem to be doing quite well. How-
ever, according to the same Figure 13, the predictions
vary for complex images with extremely diversified
sub-regions. In addition, it is clear that U-SegNet is
good at predicting regions in images that are very dif-
ficult and complex. Interestingly, U-SegNet incorpo-
rates the good features of both U-Net and SegNet ar-
chitectures. Compared to U-Net, U-SegNet has fewer
Brain Tumor Segmentation of Lower-Grade Glioma Across MRI Images Using Hybrid Convolutional Neural Networks
461