Alzheimer’s Disease Early Detection Using a Densenet-121
9
implementation general. It works with all kinds of cuboids. Also, we added a
channels parameter because the implementation always considered 3 channels
(usually red, green and blue colors), but the magnetic resonance images are
monochromatic.
Using this implementation we configure the training process of the neural
network with the following parameters.
Training
We use 75% of the data obtained from ADNI as the training
dataset. The data is obtained randomly from the complete data set
Batch size For the phase of training, we use a batch size of 5 MRI based on
experimental results by [2]
Testing
The testing dataset is the remaining 25% of the data
Channels
We send a parameter of 1 to the constructor of the neural network
because the images are monochromatic
Classes
Initially, we sent a parameter of 6 to the constructor of the neural
network. However, we decided to eliminate the SMC class because it
is a subjective class. We consider it training noise. Finally, we use a
parameter of 5 classes to classify
Dropout
We use a dropout rate of 0.7 based on observations by [2]. This
prevents overfitting
Loss
We use a cross-entropy loss function. It is useful in classification
problems that are not binary and, in our case, we have 5 or 6 classes
Optimizer We use stochastic gradient descent (SGD). This popular optimizer
is useful in the case of unbalanced data, which is our case
Learning
In the SGD optimizer, we use a learning rate parameter of 0.1 and a
drop in the learning rate in the sixty epoch of 0.1. The latter
reduces the learning rate to 0.01 in that epoch
Momentum Since the SGD optimizer with momentum usually finds flatter local
minima, we use a typical momentum of 0.9
Epochs
Since we use the Google Collaboratory platform, we set the
maximum number of epochs to 80. It was not possible to exceed
above 90 epochs to reach 100 epochs because the platform
disconnects us before achieving it
With that parameters, we pushed the limits of the Google Colaboratory
platform to produce a state-of-the-art DNN. Although other authors claim that
the free-of-charge resources of Google Colaboratory “are far from enough to
solve demanding real-world problems and are not scalable” [3], we use it as the
platform that provides us Graphics Processing Unit (GPU) computation. This
decision has limitations and implications. As explained in [3], there only 12 h
of free use of the GPU backend. We have even noticed less sometimes, approximately 10 h. After that time, Google Colaboratory disconnects and deletes the
virtual machine provided. If the user reconnects, the new machine supplied only
offers 3 h of GPU backend. After that, it is not possible to connect to a backend
9
implementation general. It works with all kinds of cuboids. Also, we added a
channels parameter because the implementation always considered 3 channels
(usually red, green and blue colors), but the magnetic resonance images are
monochromatic.
Using this implementation we configure the training process of the neural
network with the following parameters.
Training
We use 75% of the data obtained from ADNI as the training
dataset. The data is obtained randomly from the complete data set
Batch size For the phase of training, we use a batch size of 5 MRI based on
experimental results by [2]
Testing
The testing dataset is the remaining 25% of the data
Channels
We send a parameter of 1 to the constructor of the neural network
because the images are monochromatic
Classes
Initially, we sent a parameter of 6 to the constructor of the neural
network. However, we decided to eliminate the SMC class because it
is a subjective class. We consider it training noise. Finally, we use a
parameter of 5 classes to classify
Dropout
We use a dropout rate of 0.7 based on observations by [2]. This
prevents overfitting
Loss
We use a cross-entropy loss function. It is useful in classification
problems that are not binary and, in our case, we have 5 or 6 classes
Optimizer We use stochastic gradient descent (SGD). This popular optimizer
is useful in the case of unbalanced data, which is our case
Learning
In the SGD optimizer, we use a learning rate parameter of 0.1 and a
drop in the learning rate in the sixty epoch of 0.1. The latter
reduces the learning rate to 0.01 in that epoch
Momentum Since the SGD optimizer with momentum usually finds flatter local
minima, we use a typical momentum of 0.9
Epochs
Since we use the Google Collaboratory platform, we set the
maximum number of epochs to 80. It was not possible to exceed
above 90 epochs to reach 100 epochs because the platform
disconnects us before achieving it
With that parameters, we pushed the limits of the Google Colaboratory
platform to produce a state-of-the-art DNN. Although other authors claim that
the free-of-charge resources of Google Colaboratory “are far from enough to
solve demanding real-world problems and are not scalable” [3], we use it as the
platform that provides us Graphics Processing Unit (GPU) computation. This
decision has limitations and implications. As explained in [3], there only 12 h
of free use of the GPU backend. We have even noticed less sometimes, approximately 10 h. After that time, Google Colaboratory disconnects and deletes the
virtual machine provided. If the user reconnects, the new machine supplied only
offers 3 h of GPU backend. After that, it is not possible to connect to a backend