=Paper=
{{Paper
|id=Vol-2563/aics_4
|storemode=property
|title=Using Magnetic Resonance Imaging to Distinguish a Healthy Brain from a Bipolar Brain: A
Transfer Learning Approach
|pdfUrl=https://ceur-ws.org/Vol-2563/aics_4.pdf
|volume=Vol-2563
|authors=Philip Martyn,Genevieve McPhilemy,Leila Nabulsi,Fiona Martyn,Colm McDonald,Dara Cannon,Michael Schukat
|dblpUrl=https://dblp.org/rec/conf/aics/MartynMNMMCS19
}}
==Using Magnetic Resonance Imaging to Distinguish a Healthy Brain from a Bipolar Brain: A
Transfer Learning Approach==
https://ceur-ws.org/Vol-2563/aics_4.pdf
Using magnetic resonance imaging to distinguish a
healthy brain from a bipolar brain: A transfer learning
approach
Martyn, P1., McPhilemy, G2., Nabulsi, L2., Martyn, F.M.2, Hallahan, B2., McDonald,
C2., Cannon, D.M2., Schukat, M1.
1
College of Engineering and Informatics, National University of Ireland Galway,
H91 TK33 Galway, Ireland
2
Centre for Neuroimaging & Cognitive Genomics (NICOG), Clinical Neuroimaging Laborato-
ry, NCBES Galway Neuroscience Centre, College of Medicine Nursing and Health Sciences,
National University of Ireland Galway, H91 TK33 Galway, Ireland.
Abstract. Bipolar Disorder (BD) is a recurrent psychiatric condition
characterised by periods of depression and (hypo)mania, it affects more than
1% of the world’s population [1]. However, accurate diagnosis can be difficult
due to the lack of diagnostic tools available to practitioners. To address this
knowledge gap this paper aims to understand how the application of transfer
learning, in the context of machine learning techniques, can be used to improve
a diagnosis of BD.
Image detection of magnetic resonance images (MRI) was undertaken to
identify features of grey matter in BD brains in comparison to healthy controls
(HC), which may constitute a biomarker of BD. Additionally, the products of
machine learning were investigated for clinical application to efficiently aid in
clinical diagnosis by an end user, through a cloud-based application.
The transfer learning model created demonstrated at 88% accuracy the
ability to detect features present in the BD brain, not present in controls. Of
limitation to this study was the amount of MR images required to train this
model. However, this project identifies that it is possible with limited resources
to create a model which may prove useful in diagnostic settings in the future.
Keywords: Transfer Learning, Bipolar Disorder, Software Engineering
1 Introduction
Bipolar disorder (BD) is a mood disorder associated with recurrent shifts in mood
alternating between depressive and (hypo)manic episodes as well as disruptions to
cognitive function, activity levels, and sleep [1]. These functional alterations have
been associated with morphological changes in the grey and white matter structures of
the brain [2]. Robust findings of reductions of cortical thickness and volume reduction
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
2
in subcortical structures, particularly within the inferior frontal gyrus, cingulate gyrus,
middle frontal gyrus, hippocampus, amygdala, and thalamus have been demonstrated
[3]. However, while these alterations have been replicated in a number of studies,
they are not present in all [3]. A number of the aforementioned cortical and subcorti-
cal areas are involved in the cortico-limbic system, a circuit responsible for emotional
recognition, response, and regulation. Of particular note are the hippocampus, thala-
mus, amygdala, and cingulate cortex.
1.1 Related Work
Machine Learning (ML) through the use of Support Vector Machines (SVM), or
Convolutional Neural Networks (CNN) has been used extensively as a method of
diagnosing brain disorders through analysing Magnetic Resonance Imaging (MRI).
Nunes et al. [4] outline an approach using SVM to diagnose BD. A combination of
validation techniques was used to examine differences in each method. A meta-
analysis of sample-level classifiers was used which generated results of 42.26% and
59.14% for sensitivity and specificity. A Leave-One-Site-Out (LOSO) validation
method was also used which generated results of 58.67% for accuracy, while also
generating 51.99% for sensitivity and 64.85% for specificity. The final method was an
aggregate subject-level validation which generated 65.23% for accuracy, 66.02% for
sensitivity and 64.90% for specificity. The LOSO and aggregate validation also gen-
erated receiver operating characteristic curve values of 60.92% and 71.49%, respec-
tively.
Iidaka [5] developed a model to diagnose patients with autism spectrum disorder
(ASD). The model used a probabilistic neural network (PNN) to classify patients
based on certain biomarkers found in resting state functional MRI. The study used a
sample size of 312 subjects with ASD and 328 with typical development, taken from
the Autism Brain Imaging Data Exchange. The PNN was shown to have an approxi-
mately 90% accuracy rate. The study relied upon the notion that “intrinsic connectivi-
ty between subdivisions of the brain is altered in patients with ASD compared to con-
trols”, [5] which can be analysed in resting state fMRI. The process taken within that
study involved a sequence of spatially realigning volumes to the mean volume and
temporally realigning the signal within each slice of MRI to that obtained in the mid-
dle slice using Whittaker–Shannon (aka sinc) interpolation. These re-sliced volumes
were then normalized to the Montreal Neurological Institute space with a voxel size
of 3 x 3 x 3mm3. The normalized images were then spatially smoothed using a Gauss-
ian kernel. The model used in the study was a PNN, described as “a PNN (Specht,
1990) is an implementation of the kernel discriminant analysis statistical algorithm,
which is organized into a multilayered feed forward neural network to perform classi-
fication” [5]. The PNN consists of 4 fully interconnected layers, the input layer, a
pattern layer, a summation later and an output layer. The input layer has an equal
number of nodes to features and is responsible for distributing input vectors to the
pattern layer. The pattern layer takes the inputs and estimates the probability density
function, which in the Iidaka study, a Gaussian function was used. The pattern layer
outputs to nodes in the summation layer based on the different classes in the model, in
3
this case whether ASD or a control. The output layer then outputs values correspond-
ing to the most appropriate choice from the current data based on the maximum prob-
ability or Bayer’s rule. The results of the model relied on both Leave One Out Cross
Validation (LOOCV) and V-fold cross validation as the validation methods. For
LOOCV the accuracy results were 89.4% while the V-fold was 77.2%, 86.9% and
90.3% for 2-fold, 10-fold and 50-fold, respectively. An investigation into confound-
ing factors was also taken into account with similar accuracy values for groups solely
related to subjects on-medication, off-medication and subjects who exhibited head
movement while the MRI was being taken.
Transfer Learning is the process of taking a pre-trained model and refitting the
model’s final predictive layer within a CNN in order for it to be redefined to another
task. This allows for a much smaller training dataset to be used within the training
process. Hon and Khan [6] used a transfer learning method to identify AD by analys-
ing MRI. They used two pre-trained models, the VGG16 CNN and the Inception
CNN. They were re-trained using a training set taken from the Open Access Series of
Imaging Studies (OASIS) consisting of 416 subjects, made up of both AD and healthy
control (HC) subjects. The data was trimmed down to a random selection of 100 AD
and 100 HC subjects. The data was then sorted into 32 of the most informative images
from the axial plane of the 3D training images, resulting in 6400 images. The average
accuracy of these models was 74.12% for the VGG untrained, 92.3% for the trained
VGG, and 96.25% for the trained Inception model.
The basic architecture of a CNN is based around a series of stages or layers. As de-
scribed in Deep Learning by LeCun et al., [7] we see a description of a typical CNN
architecture. The initial stages of a CNN architecture are comprised of two types of
layers, these are the convolutional layers and pooling layers. The convolutional layers
are made up of feature maps. Inside of these, individual nodes connect to local patch-
es of the feature maps from the previous layer through a set of weights called a filter
bank. The sum of all these weights are then passed through a non-linearity such as a
rectified linear unit (ReLU). A ReLU is an activation function that computes f(x) =
max(0,x). In other terms it thresholds values at 0, outputting 0 when x < 0, and out-
putting a linear function when x ≥ 0 [8]. Feature maps in a layer share filter banks,
while different feature maps use different filter banks. This is due to the fact that, in
images, local motifs usually become evident which results in groupings of values in
array data. Also, these motifs are invariant to location so they could appear in any
area of the image. So, this means that individual units across the scope of image
would need to share weights to detect these individual motifs. The pooling layer in a
CNN is tasked with merging semantically similar features into one. Typically, this
will be done by computing the maximum of a local patch of units in one or multiple
feature maps. A CNN will usually be comprised of multiples of the convolutional,
non-linearity and pooling layers, followed by more convolutional and fully-connected
layers [7].
Training within a CNN happens through a process called backpropagation using
stochastic gradient descent. “The backpropagation procedure to compute the gradient
of an objective function with respect to the weights of a multilayer stack of modules is
nothing more than a practical application of the chain rule for derivatives” [7]. Back-
4
propagation computes the gradients by working backwards from the gradient with
respect to the outputs of that module, repeatedly applying this process to propagate
gradients though all modules from output to input.
Cross-validation is the process of estimating the accuracy of a prediction model. In
Iidaka [5] a description of the Leave One Out Cross validation (LOOCV) method is
shown. This method removes one subject’s data from the testing dataset and uses the
remainder for training a neural network. The removed data is then used to predict the
outcome of the removed dataset to analyse the accuracy of the model. The LOOCV
validation method is commonly used in MRI based models [9] [10]. Also in Iikada [5]
we see a description of another validation method, V-fold cross- validation. With this
validation method, a model is built with (V-1)/V proportion of the subjects used. The
remaining 1/V is then used to validate the model.
2 Methods
The output of this research involved the creation of a transfer learned convolutional
neural network (CNN), using BD and HC MRI data, as well as a deployable software
application with which to give access to the prediction capabilities of the CNN to an
end user. During the course of the transfer learning training, metrics were used to
gather data on the accuracy of the trained model. A full software development lifecy-
cle was also employed during the course of the development of the software applica-
tion. This included design, project management, development, testing and deployment
of the application.
The methodology used in the research was quantitative in nature as the essence of
the study was to produce new data to show that transfer learning is a viable option in
BD diagnostic tooling. The research carried out involved establishing accuracy data
from CNN training using transfer learning. This would give an indication as to the
veracity of the claim that transfer learning can be used in diagnosing BD using MRI
data. The process of creating an application to provide a mechanism for a user to ac-
cept the predictive capabilities of the model also produces data, but possibly of a more
abstract nature. This would include overall performance of the application in use but
also the overall user experience of the application.
2.1 Transfer Learning
The transfer learning portion of the project was achieved using the TensorFlow ma-
chine learning backend as well as the Keras library. The code for the ML part of the
project was written in Python. Keras supports multiple languages but Python is prob-
ably the most well supported language for machine learning. This meant it was a good
choice for the project. The general design of the data augmentation and model train-
ing was taken from a previous study involving transfer learning and MRI prediction
of Alzheimer’s Disease MRI data [6]. The approach taken in this study led to very
5
high accuracy, so it was deemed a good idea to follow a similar approach. This also
mirrored similar methods outlined by Francois Chollet, the author of Keras, in a Keras
tutorial [12] describing transfer learning.
A high-level view of the transfer learning process involved a step-by-step process
of data acquisition, data augmentation, training and testing. This project took a staged
approach to this process, whereby at different times different levels of data were using
in training, mainly due to difficulty in accessing a large dataset. As well as this, dif-
ferent levels of data augmentation were used to try to overcome this data problem. In
the end, the only real solution was to source more data. The results of this varying
dataset show an interesting difference in accuracy. This highlights the importance of
the size and quality of a dataset when conducting machine learning experimentation.
The different attempts made during this project to utilise the varying size and quality
datasets shows this to hold true.
The training process involved heavily leaning on the functions provided by the
Keras library, which included functions to ingest data into a format that the Tensor-
Flow engine could then use while training. Other functions involved making predic-
tions using the trained CNN as well as providing the base model for the transfer learn-
ing process. The base model used in the project was the VGG16 model, created by K.
Simonyan and A. Zisserman of Oxford University for the ImageNet competition [13].
Keras comes with this model as standard, making accessing the model very simple.
Other Python packages were also used during the machine learning portions of the
paper. These included the PILLOW and Open-CV packages for image processing,
and the Nibabel package for interacting with NII files. Med2image was used for split-
ting NII files into individual JPEG images, while the DeepBrain package was used for
extracting the brain structure from the non-brain parts of the MRI images. These
packages allowed for quicker development than rewriting the same functionality from
the ground up.
During the early phases of transfer learning research, it was necessary to format the
available data into the required directory structures in order for Keras to be able to
ingest the image data. This was performed by a combination of Bash and Python
scripts. These scripts created the directory structure, converted the MRI data to JPG,
augmented the data and saved the JPG images required for later training.
2.2 Data
Data Acquisition
The data used during the training of the CNN came from two different sources. These
were the Anatomy Department of NUI, Galway and the OpenNeuro database [14].
Originally it was hoped that the data from NUI, Galway would be sufficient but dur-
ing the course of training the CNN it became clear that this would not be the case.
Some interesting results were elicited from the earlier training sessions using just the
NUI, Galway dataset. These results showed the necessity of a sufficiently large data
6
set for machine learning. The OpenNeuro data was sourced to allow for better results
from the training.
Some issues did arise from having two separate datasets. One of these issues was
that the data was sufficiently different, visually, to necessitate different approaches to
data augmentation. The data from OpenNeuro was of a slightly lower quality and of a
different image size compared to the NUI, Galway images. This meant that the data
augmentation scripts needed to be adjusted for the differently sourced datasets.
One other attempt was made to get more data through the COINS online dataset
but the BD MRIs that were present in that dataset were of the sagittal plane and not
the axial (transverse) plane like all the other MRI. This was then deemed unusable in
the model and thus discarded.
In relation to the NUI Galway dataset, MR images were obtained from 98 individ-
uals aged 16-60 years of age as part of the Clinical Neuroimaging Laboratory Re-
search Programme. All participants provided written informed consent for the rele-
vant studies and ethical approval was obtained from the NUI Galway and Galway
University Hospitals Research Ethics Committees. As OpenNeuro was from a public
dataset all that was required to make a request to use the data and it was made availa-
ble. A further set of images was obtained from those used in the Hon & Khan [6]
study. These originally came from the Open Access Series of Imaging Studies, a pub-
licly accessible dataset.
Datasets
The NUI, Galway data was made up of 42 BD MRI and 56 HC MRI. The OpenNeuro
data consisted 49 MRI scans all from individuals with BD. This left a need for extra
HC MRI images. To fill this need some of the images from the Hon & Khan [6] study
were used to supplement the HC images used in this study. Similar pre-processing
and training strategies were utilized in the datasets of this study and the Hon & Khan
study. These included similar sizing, cropping, contrast adjustment and brain extrac-
tion and so the images used would suit for both studies. The quality of the OpenNeuro
dataset was slightly lower than that of the NUI, Galway dataset, with more severe
evidence of ghosting and distortion on the images. Most of this was towards the be-
ginning of the MR images and so was usually discarded during the image selection
procession.
3 Results
3.1 Initial Investigation
The training of the CNN can largely be broken into two phases, pre-OpenNeuro data
being added and post-OpenNeuro data being added. The results of the initial training
sessions before OpenNeuro, essentially showed how the size of the data set was not
sufficient to produce any learning in the model. The results showed overfitting in the
7
model with validation accuracy repeatedly returning ~50% and high training accura-
cy. When the resulting models were used in prediction scenarios the results were al-
ways the first class-label entered into the model when compiling, prior to training,
showing that the model was constantly picking the first class for whatever it saw eve-
ry time during training, which explains the 50% accuracy.
The sequence of initial training attempts went as follows and is summarized in Ta-
ble 1 but the general takeaway from the initial training attempts is that more data was
required for the training to be successful. All the of the initial nine attempts at training
with the initial showed similar results, hovering around ~50%. Between the attempts
variations in the training strategy were employed such as including data augmenta-
tion, changing the optimizer during training and adding further layers to the top model
of the CNN, or adding weight regularizers to these layers was also employed. None of
these strategies was successful at increasing the validation accuracy of the model.
3.2 Full Dataset
Following the last attempt at training it was clear that the amount of data was insuffi-
cient. This meant that enlarging and retraining on the larger dataset was necessary.
Following the inclusion of the OpenNeuro data, a massive improvement in the train-
ing results was elicited. After just one round of training the validation accuracy went
up to ~75%. This showed a reduction in overfitting. The validation accuracy im-
proved after the first 10-20 epochs and then stayed static hovering around 70-75%
accuracy. The validation loss also showed signs of improvement with values varying
between 0.8 and 1.2. The validation loss also increased the further into the training the
process went.
Table 1. Summary of training attempts
Training Attempt Validation Accuracy % Notes
1 ~50 Initial setup.
2 ~50-55 Changed optimiser to Adam.
Increasing levels of image pre-
processing and data augmentation to
3-9 ~50
increase size of dataset. Adjustments
to optimisers.
Added more layers to CNN top mod-
10 ~50
el with added weight regularizers.
Combined the NUI, Galway dataset
11 ~75 with OpenNeuro dataset. Large jump
in validation accuracy.
8
Following the inclusion of the OpenNeuro data, another change was made to the data.
This was to heighten the contrast of each MRI, in order to increase the detail of the
image. Also, a batch normalisation layer was included in the top model. Following
these changes and a new round of training of the image data, validation accuracy went
up to approximate average of ~77%. This configuration of was then extended to a full
5-fold cross-validation. The 5-fold cross-validation results (Table 2) show the differ-
ence in validation accuracy across the training datasets splits. The results from these
training can then be calculated to find the average total accuracy of the model of 88%
across the 5 folds. Due to some considerable threshing across the different folds, it
might have been useful to perform 10-fold cross validation. This is something which
could be taken into consideration in further future experimentation.
Tests were performed on the model produced from the final training run and the re-
sults were promising. Due to the small dataset, a decision was made to only hold back
3 BD MRIs for separate testing. Several control MRIs were also available as this type
of MRI outnumbered the BD MRIs. The results from predicting show that the majori-
ty of the selected image slices from the BD MRIs returned Bipolar as the prediction-
class, although the prediction probability value was sometimes very low and erratic.
This may be the result of some error in the process or the fact the model’s probability
of it being BD was just deemed low. Some prediction runs returned the opposite re-
sults as expected which is not a promising result. This shows some of the downsides
of the current model. Some results also returned false positives from control MRIs.
3.3 Application Performance
The performance of the application when run over the network on Amazon AWS, is
similar to the performance seen when tested locally. The main difference seen in the
time from request start to finish is the time taken to upload the MRI from the client to
the server. This time is between 5-10 seconds, depending upon the quality of the net-
work at the time. The actual processing time of a request remains the same as a local-
ly tested request at ~15 seconds.
The overall user experience of the application is a largely subjective metric but to
due to the simplicity and general stability of the program experienced during testing,
the overall user experience is generally good. The major aspects negatively impacting
user experience are the wait times during the processing of the MRI, as well as false
alarms and general inconsistent hit rates in some results. If these were reduced, user
experience would be greatly improved. The majority of the processing time is spent
on the image pre-processing and the predicting itself. This is largely due to the num-
ber of individual tasks that need to happen during these processes, including I/O and
spooling up of TensorFlow backends. These processes take time and as such means
that the frontend client may have to wait a long time. These would be interesting areas
for future research to improve the quality of the application
9
4 Discussion
The primary outcome from the results of the transfer learning training is that for pre-
dicting BD, using transfer learning is at least a semi-viable approach. The literature
suggests that certain morphology exists in the brain of BD sufferers distinct from that
of a healthy control [16] [17] and from the results seen in this paper it is evident that
transfer learning can be used to detect this morphology to a certain degree. From the
validation accuracy of 88% with the limited dataset at hand, this shows that with fur-
ther data and more time spent on improving the quality of the model or using a differ-
ent base model, higher accuracy could be achieved. This is shown by the results (Ta-
ble 3) seen in the Hon & Khan [6] study, which served as the basis for this project, the
results attained by a similar process were markedly higher. One reason for this higher
accuracy could be down to the dataset. The amount of data used by Hon and Khan in
their study was slightly larger, by roughly 10%, and their results show accuracy of
92.3% using the same VGG16 model. When they used the Inception V4 model the
average accuracy went up to 96.25%. We can see here that if there was a larger da-
taset of BD MRI then it may be possible to achieve similar levels of accuracy with a
model trained using transfer learning to detect BD. Another aspect of their data was
homogeneity. The data came from a single source and was very similarly processed
prior to ingestion into the model for training whereas the data used in this study came
from two sources and needed to be processed in different ways resulting in slightly
different images being entered into the model when training.
Table 2. Results from Hon & Khan [6] study
Model Avg. Acc. (st. dev.) (%)
VGG16 (from scratch) 74.12 (1.55)
VGG16 (transfer learning) 92.3 (2.42)
Inception V4 (transfer learning) 96.25 (1.2)
Another set of results with which to compare the results attained in this study is from
Nunes et al [4]. This study used support vector machines rather than a transfer learned
CNN for BD classification, and highlights differences between the capabilities of the
different classification methods. The results of that study used a Leave-one-site-out
(LOSO) cross-validation mechanism returning an accuracy of 58.70% and an aggre-
gate subject-level accuracy of 65.23%. This is a difference of ~30% for LOSO cross-
validation and ~23% for aggregate subject-level accuracy compared to the results
seen in this paper. This may be due to the heterogenous multisite dataset used in that
study. A homogenous dataset may improve accuracy within that single dataset, but
increased variation may lead to an overall decline in accuracy as patterns become
harder to be learned by the model.
Classification errors between the BD and HC groups during image prediction could
be related to certain confounding factors such as medication being taken by the sub-
10
jects at the time the images were taken. For instance, Lithium use has been associated
with increased thickness of the anterior cingulate cortex [18] when taken as a treat-
ment for BD. The cingulate is otherwise thinner in BD, so medication is sometimes
used to correct this [19]. This could then lead to errors in classification.
In conclusion, we have shown that transfer learning applied to feature detection in
BD does return positive results, while not being totally conclusive in its intended
outcome. In terms of future research, the project shows conclusively that further in-
vestigation into this area is warranted.
References
1. Grande, I., Berk, M., Birmaher, B., Vieta, E.: Bipolar disorder. The Lancet. 387,
1561–1572 (2016). https://doi.org/10.1016/S0140-6736(15)00241-X.
2. Hibar, D.P., Westlye, L.T., van Erp, T.G.M., Rasmussen, J., Leonardo, C.D.,
Faskowitz, J., Haukvik, U.K., Hartberg, C.B., Doan, N.T., Agartz, I., Dale, A.M.,
Gruber, O., Krämer, B., Trost, S., Liberg, B., Abé, C., Ekman, C.J., Ingvar, M.,
Landén, M., Fears, S.C., Freimer, N.B., Bearden, C.E., the Costa Rica/Colombia
Consortium for Genetic Investigation of Bipolar Endophenotypes, Sprooten, E.,
Glahn, D.C., Pearlson, G.D., Emsell, L., Kenney, J., Scanlon, C., McDonald, C.,
Cannon, D.M., Almeida, J., Versace, A., Caseras, X., Lawrence, N.S., Phillips,
M.L., Dima, D., Delvecchio, G., Frangou, S., Satterthwaite, T.D., Wolf, D.,
Houenou, J., Henry, C., Malt, U.F., Bøen, E., Elvsåshagen, T., Young, A.H.,
Lloyd, A.J., Goodwin, G.M., Mackay, C.E., Bourne, C., Bilderbeck, A.,
Abramovic, L., Boks, M.P., van Haren, N.E.M., Ophoff, R.A., Kahn, R.S., Bau-
er, M., Pfennig, A., Alda, M., Hajek, T., Mwangi, B., Soares, J.C., Nickson, T.,
Dimitrova, R., Sussmann, J.E., Hagenaars, S., Whalley, H.C., McIntosh, A.M.,
Thompson, P.M., Andreassen, O.A.: Subcortical volumetric abnormalities in bi-
polar disorder. Molecular Psychiatry. 21, 1710 (2016).
3. Hajek, T., McIntyre, R., Alda, M.: Bipolar disorders, type 2 diabetes mellitus,
and the brain: Current Opinion in Psychiatry. 29, 1–6 (2016).
https://doi.org/10.1097/YCO.0000000000000215.
4. Nunes, A., Schnack, H.G., Ching, C.R.K., Agartz, I., Akudjedu, T.N., Alda, M.,
Alnæs, D., Alonso-Lana, S., Bauer, J., Baune, B.T., Bøen, E., Bonnin, C. del M.,
Busatto, G.F., Canales-Rodríguez, E.J., Cannon, D.M., Caseras, X., Chaim-
Avancini, T.M., Dannlowski, U., Díaz-Zuluaga, A.M., Dietsche, B., Doan, N.T.,
Duchesnay, E., Elvsåshagen, T., Emden, D., Eyler, L.T., Fatjó-Vilas, M., Favre,
P., Foley, S.F., Fullerton, J.M., Glahn, D.C., Goikolea, J.M., Grotegerd, D.,
Hahn, T., Henry, C., Hibar, D.P., Houenou, J., Howells, F.M., Jahanshad, N.,
Kaufmann, T., Kenney, J., Kircher, T.T.J., Krug, A., Lagerberg, T.V., Lenroot,
R.K., López-Jaramillo, C., Machado-Vieira, R., Malt, U.F., McDonald, C.,
Mitchell, P.B., Mwangi, B., Nabulsi, L., Opel, N., Overs, B.J., Pineda-Zapata,
J.A., Pomarol-Clotet, E., Redlich, R., Roberts, G., Rosa, P.G., Salvador, R., Sat-
terthwaite, T.D., Soares, J.C., Stein, D.J., Temmingh, H.S., Trappenberg, T.,
Uhlmann, A., Haren, N.E.M. van, Vieta, E., Westlye, L.T., Wolf, D.H., Yüksel,
11
D., Zanetti, M.V., Andreassen, O.A., Thompson, P.M., Hajek, T.: Using structur-
al MRI to identify bipolar disorders – 13 site machine learning study in 3020 in-
dividuals from the ENIGMA Bipolar Disorders Working Group. Molecular Psy-
chiatry. 1 (2018). https://doi.org/10.1038/s41380-018-0228-9.
5. Iidaka, T.: Resting state functional magnetic resonance imaging and neural net-
work classified autism and control. Cortex. 63, 55–67 (2015).
https://doi.org/10.1016/j.cortex.2014.08.011.
6. Hon, M., Khan, N.M.: Towards Alzheimer’s disease classification through trans-
fer learning. In: 2017 IEEE International Conference on Bioinformatics and Bi-
omedicine (BIBM). pp. 1166–1169. IEEE, Kansas City, MO (2017).
https://doi.org/10.1109/BIBM.2017.8217822.
7. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature. 521, 436–444 (2015).
https://doi.org/10.1038/nature14539.
8. Agarap, A.F.: Deep Learning using Rectified Linear Units (ReLU).
arXiv:1803.08375 [cs, stat]. (2018).
9. Fu, C.H.Y., Mourao-Miranda, J., Costafreda, S.G., Khanna, A., Marquand, A.F.,
Williams, S.C.R., Brammer, M.J.: Pattern Classification of Sad Facial Pro-
cessing: Toward the Development of Neurobiological Markers in Depression.
Biological Psychiatry. 63, 656–662 (2008).
https://doi.org/10.1016/j.biopsych.2007.08.020.
10. Zeng, L.-L., Shen, H., Liu, L., Wang, L., Li, B., Fang, P., Zhou, Z., Li, Y., Hu,
D.: Identifying major depression using whole-brain functional connectivity: a
multivariate pattern analysis. Brain. 135, 1498–1507 (2012).
https://doi.org/10.1093/brain/aws059.
11. Sabahi, F.: Secure Virtualization for Cloud Environment Using Hypervisor-based
Technology. IJMLC. 39–45 (2012). https://doi.org/10.7763/IJMLC.2012.V2.87.
12. Chollet, F.: Building powerful image classification models using very little data,
https://blog.keras.io/building-powerful-image-classification-models-using-very-
little-data.html, last accessed 2019/08/01.
13. VGG16 - Convolutional Network for Classification and Detection,
https://neurohive.io/en/popular-networks/vgg16/, last accessed 2019/07/23.
14. OpenNeuro, https://openneuro.org/, last accessed 2019/07/28.
15. Brownlee, J.: How to Use Weight Decay to Reduce Overfitting of Neural Net-
work in Keras, https://machinelearningmastery.com/how-to-reduce-overfitting-
in-deep-learning-with-weight-regularization/, last accessed 2019/08/21.
16. Hajek, T., Cullis, J., Novak, T., Kopecek, M., Blagdon, R., Propper, L., Stopko-
va, P., Duffy, A., Hoschl, C., Uher, R., Paus, T., Young, L.T., Alda, M.: Brain
Structural Signature of Familial Predisposition for Bipolar Disorder: Replicable
Evidence For Involvement of the Right Inferior Frontal Gyrus. Biological Psy-
chiatry. 73, 144–152 (2013). https://doi.org/10.1016/j.biopsych.2012.06.015.
17. Hibar, D.P., Doan, N.T., Jahanshad, N., Cheung, J.W., Ching, C.R.K., Versace,
A., Bilderbeck, A.C., Uhlmann, A., Mwangi, B., Krämer, B., Overs, B., Hart-
berg, C.B., Abé, C., Dima, D., Grotegerd, D., Sprooten, E., Bøen, E., Jimenez,
E., Howells, F.M., Delvecchio, G., Temmingh, H., Starke, J., Almeida, J.R.C.,
Goikolea, J.M., Houenou, J., Beard, L.M., Rauer, L., Abramovic, L., Bonnin, M.,
12
Ponteduro, M.F., Keil, M., Rive, M.M., Yao, N., Yalin, N., Najt, P., Rosa, P.G.,
Redlich, R., Trost, S., Hagenaars, S., Fears, S.C., Alonso-Lana, S., van Erp,
T.G.M., Nickson, T., Chaim-Avancini, T.M., Meier, T.B., Elvsåshagen, T.,
Haukvik, U.K., Lee, W.H., Schene, A.H., Lloyd, A.J., Young, A.H., Nugent, A.,
Dale, A.M., Pfennig, A., McIntosh, A.M., Lafer, B., Baune, B.T., Ekman, C.J.,
Zarate, C.A., Bearden, C.E., Henry, C., Simhandl, C., McDonald, C., Bourne, C.,
Stein, D.J., Wolf, D.H., Cannon, D.M., Glahn, D.C., Veltman, D.J., Pomarol-
Clotet, E., Vieta, E., Canales-Rodriguez, E.J., Nery, F.G., Duran, F.L.S., Busatto,
G.F., Roberts, G., Pearlson, G.D., Goodwin, G.M., Kugel, H., Whalley, H.C.,
Ruhe, H.G., Soares, J.C., Fullerton, J.M., Rybakowski, J.K., Savitz, J., Chaim,
K.T., Fatjó-Vilas, M., Soeiro-de-Souza, M.G., Boks, M.P., Zanetti, M.V.,
Otaduy, M.C.G., Schaufelberger, M.S., Alda, M., Ingvar, M., Phillips, M.L.,
Kempton, M.J., Bauer, M., Landén, M., Lawrence, N.S., van Haren, N.E.M.,
Horn, N.R., Freimer, N.B., Gruber, O., Schofield, P.R., Mitchell, P.B., Kahn,
R.S., Lenroot, R., Machado-Vieira, R., Ophoff, R.A., Sarró, S., Frangou, S., Sat-
terthwaite, T.D., Hajek, T., Dannlowski, U., Malt, U.F., Arolt, V., Gattaz, W.F.,
Drevets, W.C., Caseras, X., Agartz, I., Thompson, P.M., Andreassen, O.A.: Cor-
tical abnormalities in bipolar disorder: an MRI analysis of 6503 individuals from
the ENIGMA Bipolar Disorder Working Group. Mol Psychiatry. 23, 932–942
(2018). https://doi.org/10.1038/mp.2017.73.
18. Emsell, L., McDonald, C.: The structural neuroimaging of bipolar disorder. In-
ternational Review of Psychiatry. 21, 297–313 (2009).
https://doi.org/10.1080/09540260902962081.
19. Machado-Vieira, R., Manji, H.K., Zarate Jr, C.A.: The role of lithium in the
treatment of bipolar disorder: convergent evidence for neurotrophic effects as a
unifying hypothesis. Bipolar Disorders. 11, 92–109 (2009).
https://doi.org/10.1111/j.1399-5618.2009.00714.x.