Page 64 - Fister jr., Iztok, Andrej Brodnik, Matjaž Krnc and Iztok Fister (eds.). StuCoSReC. Proceedings of the 2019 6th Student Computer Science Research Conference. Koper: University of Primorska Press, 2019
P. 64
lected head CT scan images called Head CT - hemor- classifying images into two target classes - images with and
rhage [8]. The dataset contains in total of 200 images of images without brain hemorrhage present.
various sizes. Of those 200 images, half of them are images
of normal head CT slides without any brain pathologies, and For the baseline experiments, we have set the parameters
the other half are the images containing some kind of brain which are we optimizing to the values presented in Table 2.
hemorrhage. Also, each image is collected from a different
person. Parameter Value
Number of neurons on the 256
last fully connected layer
0.5
Dropout probability RMSprop
Optimizer function
10−5
Learning rate
Figure 2: Example images of head CT scans, where Table 2: Baseline experiment parameter settings for
a) represents normal head CT scan image. while transfer learning fine tuning.
b) represents the head CT scan image with brain
hemorrhage present. With the presented parameter settings, we trained the CNN
for 50 epochs utilizing an efficient mini-batch training, with
batch size set to 32. As presented in the dataset section, the
collected image sizes vary from 100 x 100 px to 300 x 300
px, thus we have decided to resize all images to the VGG16
default input size of 224 x 224 px.
4.2 Grey Wolf Optimizer settings 4.4 GWOTLT settings
To initialize the GWO algorithm, tackling the problem of As presented in the previous section, the GWOTLT fine-
finding the best suitable set of parameters to achieve the tuning transfer learning parameters are set based on the
best performance of transfer learning fine-tuning, the GWO produced GWO solution. The overall architecture of the
parameter settings presented in Table 1 were used. convolutional base and the appended classification layers at
the bottom are the same as in the baseline experiment. Due
Parameter Value to the iterative nature of our proposed method, we had to
split the given train set in ratio 80:20, where we used the
Dimension of the problem 4 larger subset for training different GWOTLT produced so-
Population size 10 lutions and evaluating them - calculating the AUC on the
50 remaining smaller subset of the initial training set. In each
Number of function evaluations 0.0 run of the GWOTLT, 50 evaluations of produced possible
Lower bound 1.0 solutions are conducted, from which the best - the one with
Upper bound the highest fitness value is selected. To evaluate each solu-
tion, we train each solution for 10 epochs and then evaluate
Table 1: The initial GWO parameter settings. its performance. The selected solution is then trained for
full 50 epochs on the whole given train dataset and finally
evaluated on the given test set.
4.3 Baseline Convolutional Neural Network 4.5 Evaluation method and metrics
For the convolutional base of our proposed method, we uti- Using the described experimental setup, we conducted two
lized the VGG16 [26] CNN architecture presented in Fig- experiments, one using the CNN transfer learning approach
ure 3, pre-trained on the imagenet [6] dataset. As we can without any optimization reported as a Baseline and one uti-
observe from the figure, the VGG16 CNN is comprised of lizing the presented GWOTLT method reported as GWOTLT.
5 convolutional blocks, which together form a convolutional For each of the experiments, we obtained six performance
base. At the bottom of the convolutional base a flatten metrics: time - reported in seconds, AUC, F − 1 score, pre-
layer, two fully connected layers and one fully-connected cision, and recall, reported in percents and kappa coefficient
layer with softmax activation function forming a classifier presented as a real value on the interval between 0 and 1.
layer are chained. By default, VGG16 CNN on the input
receives an image of size 224 x 224 pixels and at the bot- To objectively evaluate the performance of the proposed
tom classifies fed images into 1000 classes, while each of the method, we adapted the gold standard 10-fold cross-validati-
convolutional layers of VGG architecture utilizes the ReLU on methodology, where a dataset is divided into train and
activation function. test sets in a ratio 90:10. Using the images from 9 out of 10
folds for the training and performing the performance eval-
Performing the transfer learning based on the VGG16 CNN uation on the remaining one fold. In the same manner, we
convolutional base, we have persisted the top four convolu- repeated the whole process in total 10 times, each time leav-
tional blocks and enabled for fine-tuning only last convolu- ing different fold out for the performance evaluation. The
tional block. At the bottom of this convolutional base, we reported values are presented as average values over 10 folds
have then chained a flatten layer, a dropout layer, fully con- if not specified otherwise.
nected layer and classifier with softmax activation function,
StuCoSReC Proceedings of the 2019 6th Student Computer Science Research Conference 64
Koper, Slovenia, 10 October
rhage [8]. The dataset contains in total of 200 images of images without brain hemorrhage present.
various sizes. Of those 200 images, half of them are images
of normal head CT slides without any brain pathologies, and For the baseline experiments, we have set the parameters
the other half are the images containing some kind of brain which are we optimizing to the values presented in Table 2.
hemorrhage. Also, each image is collected from a different
person. Parameter Value
Number of neurons on the 256
last fully connected layer
0.5
Dropout probability RMSprop
Optimizer function
10−5
Learning rate
Figure 2: Example images of head CT scans, where Table 2: Baseline experiment parameter settings for
a) represents normal head CT scan image. while transfer learning fine tuning.
b) represents the head CT scan image with brain
hemorrhage present. With the presented parameter settings, we trained the CNN
for 50 epochs utilizing an efficient mini-batch training, with
batch size set to 32. As presented in the dataset section, the
collected image sizes vary from 100 x 100 px to 300 x 300
px, thus we have decided to resize all images to the VGG16
default input size of 224 x 224 px.
4.2 Grey Wolf Optimizer settings 4.4 GWOTLT settings
To initialize the GWO algorithm, tackling the problem of As presented in the previous section, the GWOTLT fine-
finding the best suitable set of parameters to achieve the tuning transfer learning parameters are set based on the
best performance of transfer learning fine-tuning, the GWO produced GWO solution. The overall architecture of the
parameter settings presented in Table 1 were used. convolutional base and the appended classification layers at
the bottom are the same as in the baseline experiment. Due
Parameter Value to the iterative nature of our proposed method, we had to
split the given train set in ratio 80:20, where we used the
Dimension of the problem 4 larger subset for training different GWOTLT produced so-
Population size 10 lutions and evaluating them - calculating the AUC on the
50 remaining smaller subset of the initial training set. In each
Number of function evaluations 0.0 run of the GWOTLT, 50 evaluations of produced possible
Lower bound 1.0 solutions are conducted, from which the best - the one with
Upper bound the highest fitness value is selected. To evaluate each solu-
tion, we train each solution for 10 epochs and then evaluate
Table 1: The initial GWO parameter settings. its performance. The selected solution is then trained for
full 50 epochs on the whole given train dataset and finally
evaluated on the given test set.
4.3 Baseline Convolutional Neural Network 4.5 Evaluation method and metrics
For the convolutional base of our proposed method, we uti- Using the described experimental setup, we conducted two
lized the VGG16 [26] CNN architecture presented in Fig- experiments, one using the CNN transfer learning approach
ure 3, pre-trained on the imagenet [6] dataset. As we can without any optimization reported as a Baseline and one uti-
observe from the figure, the VGG16 CNN is comprised of lizing the presented GWOTLT method reported as GWOTLT.
5 convolutional blocks, which together form a convolutional For each of the experiments, we obtained six performance
base. At the bottom of the convolutional base a flatten metrics: time - reported in seconds, AUC, F − 1 score, pre-
layer, two fully connected layers and one fully-connected cision, and recall, reported in percents and kappa coefficient
layer with softmax activation function forming a classifier presented as a real value on the interval between 0 and 1.
layer are chained. By default, VGG16 CNN on the input
receives an image of size 224 x 224 pixels and at the bot- To objectively evaluate the performance of the proposed
tom classifies fed images into 1000 classes, while each of the method, we adapted the gold standard 10-fold cross-validati-
convolutional layers of VGG architecture utilizes the ReLU on methodology, where a dataset is divided into train and
activation function. test sets in a ratio 90:10. Using the images from 9 out of 10
folds for the training and performing the performance eval-
Performing the transfer learning based on the VGG16 CNN uation on the remaining one fold. In the same manner, we
convolutional base, we have persisted the top four convolu- repeated the whole process in total 10 times, each time leav-
tional blocks and enabled for fine-tuning only last convolu- ing different fold out for the performance evaluation. The
tional block. At the bottom of this convolutional base, we reported values are presented as average values over 10 folds
have then chained a flatten layer, a dropout layer, fully con- if not specified otherwise.
nected layer and classifier with softmax activation function,
StuCoSReC Proceedings of the 2019 6th Student Computer Science Research Conference 64
Koper, Slovenia, 10 October
