Deep Learning Approach Versus Traditional

Machine Learning for ADHD Classification

Gulay Cicek Aydin AKAN
Department of Biomedical Engineering Department of Electrical and Electronics Engineering
Istanbul University - Cerrahpasa Izmir University of Economics
2021 Medical Technologies Congress (TIPTEKNO) | 978-1-6654-3663-2/21/$31.00 ©2021 IEEE | DOI: 10.1109/TIPTEKNO53239.2021.9632940

Department of Software Engineering Izmir, TURKEY

Beykent University [email protected]
Istanbul, TURKEY
[email protected]

Abstract—Magnetic resonance is the imaging method that that evidence-based decisions are made. Thanks to scientific
stands out in the evaluation of textures and diseases related advances, early detection of diseases will be facilitated with
to brain. The information about metabolic, biochemical and medical imaging techniques and easier treatment opportuni-
hemodynamic structure of the brain is obtained by magnetic
resonance imaging. Attention Deficit Hyperactivity Disorder ties will be provided. Great success has been achieved in
(ADHD) is a psychiatric disease and, if not treated, its effects medical imaging with modern medical technology. Computed
may spread over all lifetime and cause significant academic, Tomography (CT), Nuclear Medicine Imaging Ultrasound, X-
social, and psychiatric problems. High-accuracy and objective Ray and Magnetic Resonance Imaging (MRI) provide detailed
tools need to be developed for classification of ADHD. In this observations of human body [6] .
study, we present machine learning (ML) and deep learning
(DL) based approaches for the classification of MR Images Deep learning, which provides intelligent decision making
collected from ADHD patients. We generate a new 2D texture by imitating the working mechanism of the human brain,
from 3-D structural magnetic resonance image by combining is a technique aimed at abstracting high-level data using
slices where gray and white matter clearly displayed. In the first deep network architectures consisting of linear and nonlinear
approach, we extract Haralick texture based features, and HOG transformations [7].
features and classify ADHD using ML methods such as Decision
Tree, K nearest neighbor, Naive Bayes, Logistic Regression, Many studies have been conducted to diagnose ADHD using
and Support Vector Machine. In the DL approach, we trained neuroimaging techniques such as Functional Magnetic Res-
four Convolutional Neural Network (CNN) structures (AlexNet, onance Imaging (fMRI), Electroencephalography, Structural
VGGNet, ResNet and GoogleNet) for ADHD classification using MRI (sMRI). The main purpose of this study is to develop an
the 2-D texture images. Classification performance obtained with objective and reliable method for the diagnosis of ADHD using
ResNet architecture in characterizing new texture is 100 %
accuracy, 100 % sensitivity, 100 % specificity. machine learning (ML), and deep learning (DL) techniques.
Studies on the application of Deep Learning Techniques for
Keywords—Attention Deficit Hyperactivity Disorder, structural diagnosing ADHD are limited.
magnetic resonance imaging, machine learning, deep learning, Jiannan et al. (2017) proposed a deep learning-based ADHD
CNN classification method. The 3-D low level features are ex-
ctracted from sMRI and fMRI data obtained from ADHD-200
I. I NTRODUCTION
Dataset. The accuracy of the proposed model is given as 69.00
Attention Deficit Hyperactivity Disorder (ADHD) is a psy- % [8].
chiatric disease and, if not treated, it affects the life quality of Zhang et al. (2017), proposed a method including sMRI
a patient causing significant academic, social and psychiatric images to develop an objective tool for ADHD diagnosis.
problems[1], [2]. Right caudate nucleus, left precuneus and left upper frontal
ADHD is characterized with symtomps of inattention, im- gyrus region and coarse segmentation were used. The sMRI
pulsivity, and hyperactivity. In the United States, the number images was smoothed with a gausssian filter. Finally, a three
of affected children increased by 5-11%, and the number of level CNN was used for classification. According to their
patient across the country increased from 2003 to 2016 by results, it was concluded that the performance rates obtained
43% [3]. with the right caudate nucleus were more succesful than other
Medical imaging routinely used in almost every medical regions[9].
branch plays an important role in the diagnosis of many Chen et al. (2019) proposed a method for the diagnosis
conditions [4]. Imaging techniques are an important part of our of ADHD by combining an EEG-based brain network with
healthcare life by assisting healthcare professionals in diag- Convolutional Neural Network (CNN). EEG information from
nosis, treatment, and prevention [5]. These techniques ensure 101 children in total was characterized by CNN-based archi-
978-1-6654-3663-2/21/ $31.00 © 2021 IEEE tectures and manual feature extraction techniques. Correlation

Authorized licensed use limited to: Univ de Alcala. Downloaded on October 19,2024 at 17:09:46 UTC from IEEE Xplore. Restrictions apply.
measurements were made between the characterized data. The
accuracy rate of the model reported as 94.67 % [10].
Zhenyu et al.(2019) proposed a method that aims to use
space-temporal deep learning models using resting state fMRI
images from the ADHD-200 dataset. In this method, various
spatial and temporal granular computation and fusion models
are proposed, including long-short-term memory (LSTM) and
spatial temporal convolution models. The accuracy rate of the
method is 71.3 %.[11].
Riaz et al. (2020), proposed a method, which includes end to
end deep learning architecture for ADHD diagnosis by using
resting state fMRI images taken from ADHD-200 Dataset. The
model called deep fMRI consist of three consecutive networks
which are feature extractor, functional connection network,
and classfication network. The accuracy rate of the proposed
method is 73.1% [12].
Magnetic Resonance images provide detailed information Fig. 1: ADHD Detection Methodology
about the anatomy and structural organization of the brain.
Some structural changes may be identified between ADHD
and control. A volume of gray and white matter and some part
of the brain especially prefrontal cortex, basal ganglia reduce.
Delays in brain maturation were observed. In this study, a
unique method has been developed for the classification of
ADHD based on the gray and white matter textural properties
that vary in many psychiatric patients.
Limited studies have been reported on the diagnosis of
ADHD using DL models. The following are the main objec-
tives and novel contributions of the present ADHD classifica-
tion study:
• We generate a new 2D texture images from 3D sMRIs
by combination of slices, and train four CNN stuctures
(AlexNet, VGGNet, ResNet, GoogleNet).
• We investigate the impact of Haralick features in the Fig. 2: Sample Brain Images
classification of ADHD model.
• We investigate the impact of HOG features in the classi-
fication of ADHD model. • Apply AlexNet, GoogleNet, ResNet and VGGNet deep
• High-level classification performance is achieved with the learning architectures for both feature extraction and
ResNet architecture used to characterize the new texture. classification on the new texture image mentioned in Step
2.
II. R ESEARCH M ETHODOLOGY • Ten Fold Cross Validation is used for evaluating the
performance of the algorithm.
Our method consists of four parts which are data acquisi- Fig.1. displays the flow chart of the proposed classifica-
tion, data extraction, classification and performance evaluation. tion methods.
The algorithm steps to be followed for our method are given
below. A. Feature Extraction and Classification
• Obtain the sMRI images from NPNeuroPsychiatric Hos- 1) Haralick Texture Feature Extraction: Characterizing the
pital and ADHD-200 Dataset. data with few attributes increase the performance of any
• Combine slices where gray and white matter appear model. In this study, Haralick feature extraction method, one
completely and clearly to create a new 2-D image. of the traditional machine learning methods, was utilized.
• Extract features with Haralick method and generate a At certain locations, the spatial dependent matrix of two
Haralick Dataset. pixels, which expresses co-formation, is created. These co-
• Extract features with HOG based features extraction occurrence matrices (GLCM) are calculated for the directions
method and generate a HOG Dataset. of 0, 45, 90, and 135 degrees [14]. Angular Second Moment,
• Classify Haralick and HOG dataset with K-Nearest Correlation, Contrast, Sum of Squares, Inverse Difference
Neighbor, Decision Tree, Logistic Regression, Naive Moment, Sum Average, Sum Variance, Sum Entropy, Entropy,
Bayes and Support Vector Machines classifiers. Difference Variance, Difference Entropy, Info. Measure of

Authorized licensed use limited to: Univ de Alcala. Downloaded on October 19,2024 at 17:09:46 UTC from IEEE Xplore. Restrictions apply.
Corr. I, Information Measure of Correlation II and Maximum E. Architecture of GoogleNet Network
Correlation Coefficient are extracted from the GLCM matrix Googlenet proposed by Szegedy et al. [21], the winner
[13] of the ILSVRC 2014 competition from Google, has been
2) Histogram of Oriented Gradients (HOG) Descriptor: selected. It achieved a top- error rate of 6.67%. This rate
One of the effective methods used in object detection problems was close to human level. Their architecture consist of 22
by detecting the shape and appearance is the histogram based layers of deep CNN. GoogleNet has 7 million parameters. This
feature extraction method. The image is divided into small architectures include 9 inception module, 4 fully connected
squares. In each square, features is extracted by calculating layer, 4 maximum pooling layer, 3 average layer, 5 fully
the derivative values in the horizontal and vertical directions, connected layer and 3 softmax layer.
and calculating the gradient and magnitude values. Then,
all features is combined in a vector, where all values in F. Architecture of ResNet Network
normalized in order to get best result[14]. ResNet won first place in the ILSVRC classification compe-
tition developed by He et al. in 2015 [15]. ResNet architecture
B. Deep Learning Networks differs from other architectures with its micro-architectural
module structure. Instead of passing through layers, it jumps
Artificial Neural Networks are formed by the connection of through some layers [16]. This situation is one of the factors
neurons by modeling the way the biological nervous system that increase success[17].
works [15]. These networks have the ability to learn, store,
and extract relationships. There are three types of layers in G. Experimental Results
neural network. These are input, hidden and output layers The most important actions to be taken for ADHD clas-
all of which contain neurons. The information is obtained sification is the detection of attributes that will characterize
when each layer performs its task. We can think of the same structural magnetic resonance images. Structural MR data
analogue as expanded in deep learning. The role of each layer consists of sections. Gray and white matter is an important
differs with the increase of layers. Deep Learning aims to parameter in the diagnosis of many psychiatric diseases. In
reach the information by providing high level abstraction with this study, a 2D texture was created by combining the sections
deep layered networks consisting of multi-level nonlinear pro- in which gray and white matter appear complete and clear.
cesses. Input data takes the data that has specified dimensions. ML and DL methods were used to characterize the formed
Convolutional layer provides to extract the new features from tissue. Haralikc and HOG methods were used as ML methods,
the data. Extracted data is reduced with the pooling layer with AlexNet, ResNet, VggNet and GoogleNet were used as DL
the given parameter. ReLU is activation layer that provides methods.
to be zero if value is negatif. In the normalization layer, the Successful classification performances were obtained with
values are drawn to certain intervals, generalization is provided all three methods.
and the error rate is reduced. The error rate is reduced and • The highest classification performance obtained with the
overfitting is removed by removing some connection between HOG dataset (Fig. 3) was obtained with the KNN classifi-
the nodes with drop out layer. Softmax layer apply some cation algorithm. The accuracy, sensitivity and specificity
probability operations in order to find the class of the data. rates of the model are % 86, % 81 and % 84, respectively.
Output layer gives the class of the data. • The decision tree was classified with the highest classi-
fication performance obtained with the Haralick dataset
C. Architecture of AlexNet Network (Fig. 5). The accuracy, sensitivity and specificity rates of
AlexNet architecture proposed by Alex Krizhevsky et al. is the model are % 99,% 99, and % 99, respectively.
• In the ADHD classification model, the highest successful
the winner of the detection and image classification competi-
tion of large-scale visual recognition(ILSVRC) in 2012 . This classification models were created with the ResNet (Fig.
architecture houses 650,000 neurons and includes 60 million 3) architecture. The accuracy, sensitivity and specificity
parameters. AlexNet architectures consist of 25 layer. There rates of the model are % 100, %100, and % 100,
is 7 ReLU, 2 normalization, 3 pooling, 2 dropout, 1 softmax, respectively.
5 convolution, 3 fully connected layer, 1 input and 1 output III. C ONCLUSION
layer.
ADHD is a neuropsychiatric disorder. As with all diseases,
early diagnosis is very important in ADHD. In such psy-
D. Architecture of VGG-16 Network
chiatric diseases, delay in diagnosis causes many psychiatric
VGGNet arhitecture invented by VGG (Visual Geometry diseases to accompany the disease, increasing the difficulty
Group) is the winner of the detection and image classifica- in diagnosis and negatively affecting every aspect of the
tion competition of large-scale visual recognition(ILSVRC) in individual’s life. The importance of neuroimaging techniques
2014. VGGNet consist of 16 fully connected layer and it is in diagnosis is undeniable. When brain MRIs of ADHD and
very attractive due to its uniform architecture. This architecture healthy individuals were compared, reductions in gray and
looks like AlexNet but has many filtres. white matter volumes were observed. In this study, gray and

Authorized licensed use limited to: Univ de Alcala. Downloaded on October 19,2024 at 17:09:46 UTC from IEEE Xplore. Restrictions apply.
 section*Conflicts of interest
All authors; Gulay Cicek and Aydin Akan declare that they
do not have conflict of interest.
F UNDING AND ACKNOWLEDGEMENT
This work was supported by Izmir Katip Celebi University
Scientific Research Projects Coordination Unit: Project num-
bers 2019-GAP-MÜMF-003 and 2017-ÖNAP-MÜMF-0002.
We would like to thank Doc.Dr. Baris Metin, Neurology
Specialist at NPIstanbul NeuroPsychiatric Hospital for his
assistance and for providing MR Data.
R EFERENCES
Fig. 3: Detailed Performance Results for Deep FeatureSet [1] S. De Silva, S. Dayarathna, G. Ariyarathne, D. Meedeniya, and S. Ja-
yarathna, “A survey of attention deficit hyperactivity disorder identifi-
cation using psychophysiological data,” International Journal of Online
and Biomedical Engineering (iJOE), vol. 15, no. 13, pp. 61–76, 2019.
[2] L. Culpepper and G. Mattingly, “Challenges in identifying and managing
attention-deficit/hyperactivity disorder in adults in the primary care
setting: a review of the literature,” Primary care companion to the
Journal of clinical psychiatry, vol. 12, no. 6, 2010.
[3] C. for Disease Control, Prevention et al., “Data and statistics about
adhd,” 2019.
[4] E. P. Balogh, B. T. Miller, and J. R. Ball, “Improving diagnosis in health
care,” 2015.
[5] D. Ganguly, S. Chakraborty, M. Balitanas, and T.-h. Kim, “Medical
imaging: A review,” in International Conference on Security-Enriched
Urban Computing and Smart Grid. Springer, 2010, pp. 504–516.
[6] K. M. Broadhouse, “The physics of mri and how we use it to reveal
the mysteries of the mind,” Different Ways to Help Those With Mental
Illness, p. 19, 2020.
[7] M. M. Najafabadi, F. Villanustre, T. M. Khoshgoftaar, N. Seliya,
Fig. 4: Detailed Performance Results for HOG FeatureSet R. Wald, and E. Muharemagic, “Deep learning applications and chal-
lenges in big data analytics,” Journal of Big Data, vol. 2, no. 1, p. 1,
2015.
[8] D. Kuang, X. Guo, X. An, Y. Zhao, and L. He, “Discrimination of
adhd based on fmri data with deep belief network,” in International
Conference on Intelligent Computing. Springer, 2014, pp. 225–232.
[9] L. Zhu, L. Zhang, Y. Han, Q. Zeng, and W. Chang, “Study of attention
deficit/hyperactivity disorder classification based on convolutional neural
networks,” Sheng wu yi xue gong cheng xue za zhi= Journal of
biomedical engineering= Shengwu yixue gongchengxue zazhi, vol. 34,
no. 1, pp. 99–105, 2017.
[10] H. Chen, Y. Song, and X. Li, “A deep learning framework for identifying
children with adhd using an eeg-based brain network,” Neurocomputing,
vol. 356, pp. 83–96, 2019.
[11] Z. Mao, Y. Su, G. Xu, X. Wang, Y. Huang, W. Yue, L. Sun, and
N. Xiong, “Spatio-temporal deep learning method for adhd fmri classi-
fication,” Information Sciences, vol. 499, pp. 1–11, 2019.
[12] A. Riaz, M. Asad, E. Alonso, and G. Slabaugh, “Deepfmri: End-to-end
deep learning for functional connectivity and classification of adhd using
Fig. 5: Detailed Performance Results for Haralick FeatureSet fmri,” Journal of Neuroscience Methods, vol. 335, p. 108506, 2020.
[13] S. F. Ershad, “Color texture classification approach based on combina-
tion of primitive pattern units and statistical features,” arXiv preprint
arXiv:1109.1133, 2011.
white matter information, which is accepted as a reference [14] X. Huang, M. Chen, and P. Liu, “Recognition of transrectal ultrasound
prostate image based on hog-lbp,” in 2019 IEEE 13th International
in many psychiatric diseases, was taken into account. For Conference on Anti-counterfeiting, Security, and Identification (ASID).
these purposes, a new 2-dimensional image was obtained from IEEE, 2019, pp. 183–187.
the combination of the sections in which these regions are [15] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
fully visible. ML and DL methods were used to characterize and pattern recognition, 2016, pp. 770–778.
this new tissue. Feature sets obtained by Haralick and HOG [16] F. DOĞAN and İ. TÜRKOĞLU, “Derin öğrenme algoritmalarinin
methods; ADHD classification models were created by giving yaprak siniflandirma başarimlarinin karşilaştirilmasi,” Sakarya Univer-
sity Journal of Computer and Information Sciences, vol. 1, no. 1, pp.
input to KNN, SVM, Decision tree, Logistic Regression classi- 10–21, 2018.
fication algorithms. ADHD classification models were created [17] E. C. Too, L. Yujian, S. Njuki, and L. Yingchun, “A comparative study
by creating automatic feature sets with AlexNet, VggNet, of fine-tuning deep learning models for plant disease identification,”
Computers and Electronics in Agriculture, vol. 161, pp. 272–279, 2019.
ResNet and GoogleNet architectures. The highest classification
performance was with ResNet architecture.

Authorized licensed use limited to: Univ de Alcala. Downloaded on October 19,2024 at 17:09:46 UTC from IEEE Xplore. Restrictions apply.