Comparing 3D Shape and Texture Descriptors Towards Tourette’s

Syndrome Prediction Using Pediatric Magnetic Resonance Imaging

Murilo Costa de Barros

1 a

, Kaue Tartarotti Nepomuceno Duarte

2 b

, Chia-Jui Hsu

Wang-Tso Lee

3 c

and Marco Antonio Garcia de Carvalho

1 d

Computing Visual Laboratory, School of Technology - UNICAMP,

R. Paschoal Marmo, 1888 - Jd. Nova It

alia, 13484-332 - Limeira, S

ao Paulo, Brazil

Vascular Imaging Laboratory, Calgary University,2500 University Dr NW, Calgary, AB T2N 1N4, Canada

Department of Pediatrics, National Taiwan University Hospital Hsinchu Branch, Hsinchu, 300001, Taiwan

Department of Pediatrics, National Taiwan University Children’s Hospital, Taiwan

Keywords:

Tourette Syndrome, GLCM, Texture Feature, Shape Feature, Image Processing, Classiﬁcation, Machine

Learning, MRI.

Abstract:

Tourette Syndrome (TS) is a neuropsychiatric disorder characterized by the presence of involuntary motor and

vocal tics, with its etiology suggesting a strong and complex genetic basis. The detection of TS is mainly

performed clinically, but brain imaging provides additional insights about anatomical structures. Interpreting

brain patterns is challenging due to the complexity of the texture and shape of the anatomical regions. This

study compares three-dimensional texture and shape features using Gray-Level Co-occurrence Matrix and

Scale-Invariant Heat Kernel Signature. These features are analyzed in the context of TS classiﬁcation (via

Support Vector Machines), focusing on anatomical regions believed to be associated with TS. The evaluation

is performed on structural Magnetic Resonance (MR) images of 68 individuals (34 TS patients and 34 healthy

subjects). Results show that shape features achieve 92.6% accuracy in brain regions like the right thalamus

and accumbens area, while texture features reach 73.5% accuracy in regions such as right putamen and left

thalamus. Majority voting ensembles using shape features obtain 96% accuracy, with texture features achiev-

ing 79.4%. These ﬁndings highlight the inﬂuence of subcortical regions in the limbic system, consistent with

existing literature on TS.

1 INTRODUCTION

Tourette Syndrome (TS) is a genetic condition with

neuroanatomical and neurophysiological alterations,

mainly characterized by motor and vocal manifesta-

tions commonly known as tics (Johnson et al., 2023).

The development of TS typically occurs before the

age of 18, with the ﬁrst symptoms potentially emerg-

ing between the ages of 4 and 6 years. The tics can

change, either by decreasing or increasing in intensity

during the growth process. TS is known to affect ap-

proximately 1% of the world population and is more

commonly diagnosed in male children (Jones et al.,

2023). Tics can be classiﬁed into two categories: sim-

https://orcid.org/0000-0003-2452-8128

https://orcid.org/0000-0002-4074-3672

https://orcid.org/0000-0003-3231-9764

https://orcid.org/0000-0002-6303-5564

ple tics, consisting of blinking, involuntary shoulder

movement, neck and lip twitching, and in some cases,

even coughing or grunting; and complex tics, which

tend to be more drastic, such as self-injurious behav-

ior or jumping (Cen et al., 2020). Diagnosing the

syndrome is challenging due to the potential presence

of other comorbidities, such as Attention Deﬁcit Hy-

peractivity Disorder (ADHD), Obsessive-Compulsive

Disorder (OCD), as well as the coexistence of anxi-

ety, depression, and other psychiatric manifestations

(Jones et al., 2023). (Figure 1)

TS is a complex topic of study due to changes in

speciﬁc brain regions, such as caudate, thalamus, and

certain cortical areas. However, these alterations are

often difﬁcult to identify solely through medical ap-

pointments. While current literature express the ad-

vantages of diagnosing TS via clinical data (Pring-

sheim et al., 2023), the increased use of medical

imaging modalities, such as structural Magnetic Res-

474

Costa de Barros, M., Duarte, K., Hsu, C., Lee, W. and Garcia de Carvalho, M.

Comparing 3D Shape and Texture Descriptors Towards Tourette’s Syndrome Prediction Using Pediatric Magnetic Resonance Imaging.

DOI: 10.5220/0012374200003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 3: VISAPP, pages

474-481

ISBN: 978-989-758-679-8; ISSN: 2184-4321

Tourette’s Syndrome

iceberg

Vocal tics

Motor tics

Attention

Deficit/Hyperactivit

y Disorder (ADHD)

Learning

Disabilities

Obsessive

Compulsive

Behaviors

Handwriting

Difficulties

Behavioral

Issues

Sleep

Issues

Impulsivity

Anxiety

Transition

Issues

Executive

Functioning

Deficits

Sensory

Processing

Issuesc

Social

Communicatin

Deficits

Figure 1: Tourette’s Syndrome: What we can see VS what

it can hide. Adapted from (Tourette.org, 2017).

onance (MR) images, is expected to provide beneﬁ-

cial outcomes since it can identify patterns in the brain

that are hidden from human eyes. These patterns can

be captured by feature descriptors, techniques that ex-

tract speciﬁc information from a given structure. Two

prominent styles of feature extraction are either by

texture (repetitive occurrence of intensity tone in a re-

gion) or by shape (complex boundaries that surround

a certain region). However, there is no consensus on

which feature extraction style is better, leaving the

choice to the speciﬁc application.

In this work, we adopted a 3D Gray-Level Co-

occurrence Matrix (GLCM) (Haralick et al., 1973)

to extract texture features, whereas the 3D Scale-

Invariant Heat Kernel Signature (SIHKS) (Kamara-

jugadda and Pavani, 2022) has been used to extract

shape features. A further processing step is often per-

form to interpret these features, which can serve as

input to classify, for example, whether a child has TS.

Machine Learning (ML) techniques are commonly

used for this purpose, as they allows for unsupervised

feature classiﬁcation.

The main goal of this work is to classify TS pa-

tients by comparing texture and shape features ob-

tained in MR images. The main contributions of this

work are: (1) classifying TS patients using shape and

texture features; (2) identifying brain regions that are

more susceptible to TS; and (3) studying which fea-

ture extractor ensemble is efﬁcient for classifying TS.

The remaining sections of this paper are organized

as follows: the Section 2 outlines the existing work

and identiﬁes the research gaps; In Section 3, it is de-

scribed our materials and proposed method; Section

4 provides our results, which are further discussed in

Section 5; Finally, Section 6 provides a summary and

conclusion of our work.

2 LITERATURE REVIEW

Adopting advanced computer methods to interpret TS

poses challenges due to slight known changes in both

texture and shape of anatomical regions. Most of

work often focus on clinical information, which might

not consider important unseen aspects of the subject’s

brain. However, such work that employs shape and

texture features have been frequently used in other

disorders (Betrouni et al., 2021).

Texture features have strong spatial characteriza-

tion that allows distinguishing between healthy and

symptomatic patients. In Silva (Silva et al., 2023),

a technique that involves GLCM and Wavelet trans-

forms is used to classify Alzheimer’s Disease (AD) in

three possible prodromal stages. They studied the im-

plication of using distinct classiﬁers, such as Support

Vector Machines (SVM) and Random Forests (RF),

towards improving the accuracy of the models. They

have shown that aligning texture and machine learn-

ing classiﬁcation has the potential to unveil undiscov-

ered brain patterns. Aside from AD, texture features

have also being adopted in the context of Parkinson’s

Disease (PD) (Betrouni et al., 2021), where the au-

thors classiﬁed two symptomatic stages (early and

late PD) against healthy normal. First- and second-

order texture features have shown that regions like

putamen, thalamus, and caudate had shown signiﬁ-

cant results. Shifting the attention to TS, 3D texture

patterns have been employed by (Barros et al., 2022)

in the context of identifying the statistically relevant

anatomical regions to identify TS. Each combination

of texture feature and anatomical regions was cor-

rected using the False-Discovery Rate (FDR), which

later indicated that regions such as putamen, caudate,

cingulate cortex, prefrontal, temporal, and parietal

cortex had the highest correlation with deﬁning TS

patients.

Shape features are also a powerful domain where

techniques are responsible to distinguish the bor-

der/shape of an object. In Yeh (Yeh, 2020), the author

adopted shape analysis to diagnose brain tumors by

examining the morphology of association pathways in

the human brain. The study not only used structural

MR images but also diffusion MR scans. However,

the literature involving shape features combined with

medical images is still an open subject, but it deserves

attention since it has been proved that it is essential to

detect changes in boundaries. As expected, no other

work in our literature showed the use of shape extrac-

tion to identify TS patients.

It has been evident that using machine learning af-

ter image features (i.e., texture and shape) improves

the classiﬁcation results. However, there is an ap-

Comparing 3D Shape and Texture Descriptors Towards Tourette’s Syndrome Prediction Using Pediatric Magnetic Resonance Imaging

475

parent limitation when those topics are implemented

towards identifying TS since most of the studies are

mainly performed using clinical input. We identiﬁed

that the major reasons for such limitation is the lack

of publicly available dataset and slight changes in the

brain between TS and healthy subjects rather than the

importance of image features.

This work differs from the literature due to: (1)

Addressing the understanding of TS using texture

and shape features; (2) Applying feature extraction

aligned with a classiﬁer in speciﬁc regions, instead

of addressing the entire brain volume; and (3) Pre-

senting a selection of region-based classiﬁers that are

more suitable to predict TS.

3 MATERIALS AND METHOD

In this section, we deﬁne our materials and method

used for this work (Figure 2). In essence, our method

consists of four main steps: (1) Data acquisition,

describing the dataset used; (2) Preprocessing and

Volume Parcellation, consisting in the preprocessing

steps followed by the partitioning of the brain into

anatomical regions; (3) Feature extraction, consider-

ing texture and shape feature extraction and (4) Clas-

siﬁcation, involving a classiﬁer to distinguish TS from

NH subjects.

Figure 2: Proposed method to classify TS patients.

3.1 Data Acquisition

For data collection, all patients followed a screening

process to select candidates based solely on ages be-

tween 6 to 14 years. The inclusion criteria of par-

ticipants are: 1) treatment naive participants with-

out other underying neurological disease (e.g., ADHD

and OCD); 2) images without excessive head motion

(artifacts) during scanning.

A total of 68 subjects were selected, organized

into two equal groups (i.e., 34 in each). The TS

group consisted of 23 males and 11 females, aged

between 6 and 13, whereas the Normal Health (H)

group comprised 24 males and 10 females, aged 6 to

14. For each patient, a structural Magnetic Resonance

(MR) images weighted in T1 (T1-w) were acquired

using the SIEMENS Triotrim model scanner. The fol-

lowing MR protocols were used for the acquisition:

time of echo = 2.98ms, repetition time = 2000ms,

inversion time = 900ms. Each volume consisted of

192 × 208 × 256, isotropic resolution of 1mm

, ﬂip

angle = 9

◦

, and slice thickness = 1mm.

3.2 Preprocessing and Volume

Parcellation

parcellate automatic the brain volume. We adopted

Freesurfer version 6.0 to automatically segment brain

volume. In general aspects, Freesurfer can be decom-

posed into three major steps, as described by Fischl

(Fischl et al., 2002). (1) Image preprocessing, in-

volving normalization, contrast enhancement, noise

reduction, and registration. The preprocessed images

are skull-stripped to retain only brain tissue (Figure

3.b), (2) Brain surface inﬂation, building a surface

model to represent the cortical area, which is then

inﬂated to registered to a spherical atlas. (3) Brain

segmentation, assigning voxel-wise labels from dis-

tinct atlases. In this work, we adopted the DKT-atlas

(Figure 3.c), which decomposes the cortical area into

smaller groups. The regions of interest (ROI) used for

our work are presented in Table 1.

(a) (b) (c)

Figure 3: Preprocessing and volume segmentation for

freesurfer. (a) Original image normalization; (b) skull strip-

ping; (c) DKT-Atlas.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

476

Table 1: Anatomical brain regions used in this work.

Deep Gray-matter Cortical area

Thalamus (L+R) Medial Orbitofrontal (L+R)

Caudate (L+R) Inferior Parietal (L+R)

Putamen (L+R) Superior Parietal (L+R)

Accumbens(L+R) Lateral Orbitofrontal (L)

Ventral DC (L+R)

L = left hemisphere, R = right hemisphere, DC = diencephalon.

3.3 Feature Extraction

Two styles of feature extraction techniques were con-

sider to this work: (1) texture, and (2) shape. Texture

and shape features are often used in controversial as-

pects, but we believe that important, yet complemen-

tary, information can be acquired from both of them.

Texture Features. The Gray Level Co-occurrence

Matrix (GLCM) was chosen to extract texture fea-

tures as it enables the analysis of the conﬁguration

of pixel transitions in a volume. This technique has

been proven to provide reliable outcomes to distin-

guish pathological/disorder stages from normal sub-

jects (Syed et al., 2023; Gengec¸ Benli and Andac¸,

2023). In summary, GLCM involves the extraction

of dependable texture features through the occurrence

matrices (Haralick et al., 1973). The gray-scale values

within the image are represented in a co-occurrence

matrix as indices of a given row v

and column v

and the interaction function denoted as f (v

, v

) sig-

niﬁes the count of occurrences from v

to v

The computation of pixel transitions primarily re-

lies on two parameters: 1) the angle θ, set to four pos-

sible directions (0

◦

, 45

◦

, 90

◦

, 135

◦

), representing the

orientation between v

and v

and stored in distinct

GLCMs; 2) the distance d, set to 1 (adjacent corre-

spondences), corresponding to the pixel or voxel in-

terval between v

and v

. Two modiﬁcations have

been introduced to the GLCM methodology: Since

3D structures are being used (26-connectivity) as op-

posed to 2D structures (8-connectivity), additional di-

rections have been incorporated to calculate the dis-

placement vector (dx, dy, dz). The GLCM analy-

sis has been reformulated to focus on anatomical re-

gions rather than the entire brain, disregarding occur-

rences of voxels that do not belong to a speciﬁc re-

gion. In essence, each anatomical region is associ-

ated with its distinct GLCM. In conclusion, to con-

struct the ﬁnal vector encapsulating information for

each region per subject, twenty-four GLCM-based

metrics have been computed: Autocorrelation; Joint

Average; Cluster Prominence; Cluster Shade; Cluster

Tendency; Contrast; Correlation; Difference Average;

Difference Entropy; Difference Variance; Joint En-

ergy; Joint Entropy; Informational Measure of Cor-

relation 1 and 2; Inverse Difference Moment; Maxi-

mal Correlation Coefﬁcient; Inverse Difference Mo-

ment Normalized; Inverse Difference; Inverse Differ-

ence Normalized; Inverse Variance; Maximum Prob-

ability; Sum Average; Sum Entropy; Sum of Squares

(pyradiomics community, 2016).

Shape Features. In this work we apply Heat Kernel

Signature (HKS) according to (Bronstein and Kokki-

nos, 2010a) studies and evaluate the changes in a

more signiﬁcant way between brain regions. HKS is

a local shape analysis that uses geometric properties

of forms(Bronstein and Kokkinos, 2010a). Through

a multiple mesh, the application of the heat diffusion

property can be represented by Equation 1.



∆ +

∂

∂t



µ(x,t) = 0 (1)

where the negative Laplace-Beltrami operator is rep-

resented by ∆, and the distribution and composition

of heat points at a location x over time t is denoted by

µ(x,t).

The main diagonal of the heat kernel is used as a

feature descriptor in surface analyses. For each point

x on the surface, an individual feature vector is com-

puted, representing its Heat Kernel Spectrum (Sun

et al., 2009). This calculation is represented by Equa-

tion 2, which describes the quantiﬁcation of heat at a

speciﬁc point x after a particular time period t.

HKS(x) = c(x)(K

(x, x), ..., K

(x, x) (2)

where the variable c(x) satisﬁes the condition

||HKS(x)||

= 1. The calculation of HKS is based on

the analysis of the eigenvalues and the initial eigen-

functions of the Laplace-Beltrami operator.

Some limitations of HKS are related to scale sen-

sitivity. In this work, to address this potential obsta-

cle, we utilize the Scale-Invariant Heat Kernel Sig-

nature (SIHKS) (Bronstein and Kokkinos, 2010b), as

described below.

diff

(x) =



logK

τ2

(x, x) − log K

τ1

(x, x),

logK

τm

(x, x) − log K

τm−1

(x, x)



(3)

SIHKS(x) = ∥(F

diff

(x))(ω

, . . . , ω

)∥ (4)

where the representation of the Fourier transform is

given by F, and the frequency is ω.

Figure 4 illustrates the decomposition of SIHKS.

Comparing 3D Shape and Texture Descriptors Towards Tourette’s Syndrome Prediction Using Pediatric Magnetic Resonance Imaging

477

Figure 4: Distinct frequencies of SIHKS applied to the su-

perior parietal cortex.

3.4 Classiﬁcation and Ensemble

Classiﬁcation. The Support Vector Machine (SVM)

was chosen due to its ability to classify linearly sepa-

rable data with two classes (TS and control) through

the construction of hyperplanes (Feng et al., 2016).

We utilized a linear kernel and set the regulariza-

tion parameter C = 1, which essentially determines

the conﬁguration of the hyperplane margin. The con-

struction of the hyperplanes can be deﬁned as:

min

w,b

w +C

∑

i=1

max(0, 1 − y

φ(x

) + b)) (4)

where, w, b are hyperplane parameters; C is the reg-

ularization parameter (strictly positive), and x repre-

sents the input variables.

After proper construction and adjustment of the

hyperplane, the SVM needs to distinguish between

the types of samples to be analyzed, as represented

by Equation 5.

I(x) = sgn(

∑

i=1

(x · x

) + b) (5)

To maintain the reliability of the results, we

adopted the application of k-fold cross validation.

The process was divided into n = 68, representing

the total number of samples in our dataset, with k =

10 indicating the partition of the data. This division

resulted in 8 sets with 7 samples each and 2 sets with

6 samples each, which were trained in 10 iterations.

Majority Voting. We adopted the ensemble via ma-

jority voting to combine the best anatomical regions

(identiﬁed by accuracy of the models). This step

is performed to aggregate predictions from multiple

models and enhance performance. Notably, we per-

formed the ensemble for two scenarios: texture fea-

tures followed by SVM and shape features followed

by SVM. The majority voting can be obtained by the

following formula:

Figure 5: Generation of a Election-based Majority voting

considering the top 3 regions in accuracy. This concept can

be extend to Top 5 regions. Predicting 1 means having TS,

and 0 is considered as normal control.

(

1 if

∑

i=1

> 0.5

0 otherwise

(6)

where R

denotes the ﬁnal response, R

is the response

per region, and N is the number of regions speciﬁed

by Top N.

In other words, the majority voting process selects

the most frequently predicted class among the ensem-

ble of regions. For instance, if, for each patient, across

the top three regions, two indicate that the patient be-

longs to the TS class, and the other indicates the nor-

mal class, then the majority vote would assign the TS

class as the ﬁnal prediction. Our results were ana-

lyzed by assessing the responses of the top three (Top

3) and top ﬁve (Top 5) best-performing regions in this

study. A more in-depth visualization of the Top 3 out-

comes is shown in Figure 5.

3.5 Computational Tools

The implementation of this work was carried out on a

macOS High Sierra version 10.13.6. with i5 process-

ing and 16Gb of RAM. Our codes was mostly devel-

oped in Python 3.8 using the Spyder platform. Our

approach was executed utilizing the subsequent tools

and methodologies:

1. Preprocessing and Volume Parcellation. Bash

scripts were created to automatically utilize the

Freesurfer tool with the recon-all command. We

separated the regions into independent volumes

(for texture analysis) and meshes (for shape anal-

ysis - using ITK).

2. Feature Extraction. We used pyRadiomics library

in Python to compute the GLCM for the volumes,

whereas we adopted an independent code to com-

pute the SIHKS in Matlab R2018a.

3. Classifying. The SVM was computed using the

scikit-learn package, while the Majority Voting

was coded from scratch.

The analyses and applications used in this study,

as well as the models and ﬁgures, can be veriﬁed and

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

478

downloaded from our freely-available GitHub reposi-

tory

4 RESULTS

In this study, ﬁve metrics were adopted, namely: (1)

Precision, denoted as P; (2) Recall, referred as R; (3)

F-score, denoted as F; (4) Accuracy, represented by

These metrics are derived by incorporating values

such as true-positive (TP), true-negative (TN), false-

positive (FP), and false-negative (FN). The calcula-

tion of these evaluation metrics is expressed by the

following formulas:

P =

T P

T P + FP

, R =

T P

T P + FN

, F = 2 ×

P ∗ R

P + R

(7)

A =

T P + T N

T P + T N + FP + FN

(8)

Each metric was applied to evaluate both individ-

ual anatomical regions and their corresponding en-

sembles. We preferred ranking the regions by accu-

racy values, and including the top three (Top 3) and

ﬁve (Top 5) best regions as part of the ensembles us-

ing Majority Voting. The best regions, along with the

ensembles, are shown both for the shape- (Table 2)

and texture-based classiﬁcation (Table 3). In Figure

6, we show the accuracy of each anatomical region,

followed by the ensembles (Top 3 and Top 5).

5 DISCUSSION

Tourette’s syndrome is an area that requires attention

since patients around the globe suffer from involun-

tary movements that can directly affect their social

integration. Patients who suffer from TS sometimes

try to mask their tics, possibly causing discomfort

and worsening the tic when released (Eapen et al.,

Table 2: Evaluation of the classiﬁcation using Shape fea-

tures.

# Region Name A P R F

Individual brain regions

1 Accumbens (R) 92.6 91.4 94.1 92.7

2 Thalamus (R) 92.6 93,9 91.1 92.5

3 Lateral orbitofrontal (L) 91.1 93.7 88.2 90.9

4 Accumbens (L) 89.7 86.4 94.1 90.1

5 Medial orbitofrontal (R) 89.7 88.5 91.1 89.8

Ensembles

Top 3 90.0 83.0 100.0 91.0

Top 5 96.0 92.0 100.0 96.0

https://github.com/hidden-for-peer-review

2016). TS is often detected clinically, but patients can

mask the tic, thus generating possible false-negatives.

However, implementing techniques to identify brain

biomarkers for TS is a challenging task due to its sup-

posedly modest brain changes. Thus, our primary ef-

fort in this paper is classifying TS based on image fea-

tures to later enable a more in-depth understanding of

brain changes.

According to Figure 6 and Table 2, notably some

crucial regions achieved higher accuracy. The right

thalamus (92.6% accuracy in shape analysis) has the

role of integrating sensory (including emotional re-

sponses) and motor stimuli between the central and

peripheral nervous systems. This result aligns with

the current literature that highlights the thalamus as a

target structure to understand TS (Baldermann et al.,

2021). The right accumbens area (92.6% accuracy in

shape analysis) plays a vital role in regulating emo-

tional stimuli and is associated with the learning pro-

cess. This region is often used to study comorbidities

frequently observed in Tourette Syndrome, such as

ADHD (McCairn et al., 2016; Zhu et al., 2016). Al-

though the texture feature results have achieved lower

accuracy when compared to shape feature results, it is

crucial to highlight the pivotal role that understand-

ing texture changes helps to better discriminate TS

from normal control. The right putamen (73.5% ac-

curacy in texture analysis) plays a role in regulating

movement, initiating and modulating voluntary body

actions. This region, which is located in a subcortical

area forming part of the striatum, has been focused

on studying TS and its patterns (Rae et al., 2019).

The thalamus located in the left hemisphere (73.5%

accuracy in texture analysis) has a vital role in relay-

ing sensory information between the right part of the

body to the cerebral cortex. This region was studied

in order to characterize TS(M

uller-Vahl et al., 2014).

Both classiﬁers (shape and texture) displayed sim-

ilar behavior in achieving high accuracy of thalamus

and the orbitofrontal and prefrontal cortex. Another

relevant pattern to mention is the location of the best

regions. In both techniques, the most inﬂuential re-

gions are subcortical, more precisely part of the lim-

Table 3: Evaluation of the classiﬁcation using Texture fea-

tures.

# Region Name A P R F

Individual brain regions

1 Putamen (R) 73.5 72.2 76.4 74.2

2 Thalamus (L) 73.5 75 70.5 72.7

3 VentralDC (R) 70.5 68.4 76.4 72.2

4 Medial orbitofrontal (R) 66.1 65.7 67.4 66.6

5 Lateral orbitofrontal (L) 66.1 67.7 61.7 64.6

Ensembles

Top 3 77.9 77.1 79.4 78.2

Top 5 79.4 77.7 82.3 79.9

Comparing 3D Shape and Texture Descriptors Towards Tourette’s Syndrome Prediction Using Pediatric Magnetic Resonance Imaging

479

Figure 6: Accuracy values by regions, among shape features, texture, and combinations of the three and ﬁve best regions.

bic system of the brain, which are stimulated during

the occurrence of tics (Leisman and Sheldon, 2022).

In order to reinforce the robustness and achieve

higher accuracy of our method, we adopted ensem-

bles, which combined the best regions in each case.

When using shape features, we achieved the high-

est performance, reaching 96% accuracy in the Top

5, whereas we achieved 79% accuracy for the Top 5

in our texture analysis. This corroborates the effec-

tiveness of using ensemble methods when combining

distinct regions to achieve higher predictiveness. This

can possibly suggest that evaluating a combination

of regions instead of single regions can be a path to

thrive in the better understanding of TS.

6 SUMMARY & CONCLUSIONS

Exploring the application of texture and shape fea-

tures to evaluate anatomical brain regions in the

context of TS classiﬁcation represents a novel area,

largely unexplored by the current state-of-the-art.

Thus, we proposed a comparative analysis between

two types of descriptors to discern which is more ef-

fective in predicting TS. We analyzed each individual

brain region separately, as well as their combination

into ensembles. We noticed that SVM demonstrated

higher accuracy when using shape features, while fur-

ther investigation is needed to grasp the potential of

texture features.

We observed that the regions of the limbic system,

more speciﬁcally the accumbens area and the thala-

mus, showed remarkable accuracy values in the anal-

ysis of shape features, while the putamen and the tha-

lamus presented higher accuracy in our texture anal-

ysis. The presence of the thalamus in both analyses

suggests that the region plays a fundamental role in

the occurrence of symptoms associated with TS. This

structure is responsible for integrating sensory infor-

mation and communicating with associative cortical

areas. On the other hand, the orbitofrontal medial cor-

tex has shown relevant accuracy in both descriptors,

aligning with the literature since it processes emo-

tional information, makes decisions, and regulates so-

cial behavior. The presence of these regions only re-

inforces the accuracy of the models presented in this

work and enables further questions about the impor-

tance of other regions.

The study sheds light to some future directions:

(1) the need for improved and more comprehensive

descriptors to distinguish TS from NH subjects; (2)

a study exploring deep learning as an automatic fea-

ture extraction method; (3) the adoption of alternative

classiﬁers and feature descriptors.

REFERENCES

Baldermann, J. C., Kuhn, J., Sch

uller, T., Kohl, S., Andrade,

P., Schleyken, S., Prinz-Langenohl, R., Hellmich, M.,

Barbe, M., Timmermann, L., Visser-Vandewalle, V.,

and Huys, D. (2021). Thalamic deep brain stimulation

for tourette syndrome: A naturalistic trial with brief

randomized, double-blinded sham-controlled periods.

Brain Stimulation, 14.

Barros, M. C., Duarte, K. T., Lee, W.-T., Hsu, C.-J., and

de Carvalho, M. A. G. (2022). Analysis of the sta-

tistical signiﬁcance of 3d texture features in mri im-

ages toward the detection of tourette’s syndrome. In

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

480

2022 35th SIBGRAPI Conference on Graphics, Pat-

terns and Images (SIBGRAPI), volume 1, pages 103–

108.

Betrouni, N., Moreau, C., Rolland, A.-S., Carriere, N.,

Chupin, M., Kuchcinski, G., Lopes, R., Viard, R.,

Defebvre, L., and Devos, D. (2021). Texture-based

markers from structural imaging correlate with motor

handicap in parkinson’s disease. Scientiﬁc Reports,

11:2724.

Bronstein, M. M. and Kokkinos, I. (2010a). Scale-invariant

heat kernel signatures for non-rigid shape recognition.

In 2010 IEEE Computer Society Conference on Com-

puter Vision and Pattern Recognition, pages 1704–

1711.

Bronstein, M. M. and Kokkinos, I. (2010b). Scale-invariant

heat kernel signatures for non-rigid shape recognition.

In 2010 IEEE Computer Society Conference on Com-

puter Vision and Pattern Recognition, pages 1704–

1711.

Cen, S.-s., Yu, J., Wang, Q., Deeb, W., Wang, K.-l., Shukla,

A. W., Malaty, I., Ramirez-Zamora, A., Zhang, J.-g.,

Hu, W., and Meng, F.-g. (2020). Multidisciplinary

telemedicine care for tourette syndrome: Minireview.

Frontiers in Neurology, 11.

Eapen, V., Cavanna, A. E., and Robertson, M. M. (2016).

Comorbidities, social impact, and quality of life in

tourette syndrome. Frontiers in psychiatry, 7:97.

Feng, W., Sun, J., Zhang, L., Cao, C., and Yang, Q. (2016).

A support vector machine based naive bayes algo-

rithm for spam ﬁltering. In 2016 IEEE 35th Interna-

tional Performance Computing and Communications

Conference (IPCCC), pages 1–8. IEEE.

Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich,

M., Haselgrove, C., van der Kouwe, A., Killiany, R.,

Kennedy, D., Klaveness, S., Montillo, A., Makris, N.,

Rosen, B., and Dale, A. M. (2002). Whole brain

segmentation: Automated labeling of neuroanatomi-

cal structures in the human brain. Neuron, 33(3):341–

355.

Gengec¸ Benli, S¸. and Andac¸, M. (2023). Construct-

ing the schizophrenia recognition method employing

glcm features from multiple brain regions and ma-

chine learning techniques. Diagnostics, 13(13):2140.

Haralick, R. M., Shanmugam, K., et al. (1973). Textural

features for image classiﬁcation. IEEE Transactions

on systems, man, and cybernetics, 6:610–621.

Johnson, K. A., Worbe, Y., Foote, K. D., Butson, C. R.,

Gunduz, A., and Okun, M. S. (2023). Tourette syn-

drome: clinical features, pathophysiology, and treat-

ment. The Lancet Neurology, 22(2):147–158.

Jones, K. S., Saylam, E., and Ramphul, K. (2023). Tourette

syndrome and other tic disorders. StatPearls.

Kamarajugadda, K. K. and Pavani, M. (2022). Multi-

features assisted age invariant face recognition and

retrieval using cnn with scale invariant heat kernel

signature. In Fernandez, M. A. A. and Travieso-

Gonzalez, C. M., editors, Artiﬁcial Intelligence An-

nual Volume 2022, chapter 8. IntechOpen, Rijeka.

Leisman, G. and Sheldon, D. (2022). Tics and emotions.

Brain Sciences, 12(2).

McCairn, K., Nagai, Y., Hori, Y., Ninomiya, T., Kikuchi,

E., Lee, J.-Y., Suhara, T., Iriki, A., Minamimoto, T.,

Takada, M., Isoda, M., and Matsumoto, M. (2016). A

primary role for nucleus accumbens and related limbic

network in vocal tics. Neuron, 89(2):300–307.

uller-Vahl, K. R., Grosskreutz, J., Prell, T., Kaufmann,

J., Bodammer, N., and Peschel, T. (2014). Tics are

caused by alterations in prefrontal areas, thalamus and

putamen, while changes in the cingulate gyrus reﬂect

secondary compensatory mechanisms. BMC Neuro-

science, 15(1).

Pringsheim, T., Ganos, C., Nilles, C., Cavanna, A. E.,

Gilbert, D. L., Greenberg, E., Hartmann, A., Hed-

derly, T., Heyman, I., Liang, H., Malaty, I., Malik,

O., Debes, N. M., Vahl, K. M., Munchau, A., Mur-

phy, T., Nagy, P., Owen, T., Rizzo, R., Skov, L., Stern,

J., Szejko, N., Worbe, Y., and Martino, D. (2023).

European society for the study of tourette syndrome

2022 criteria for clinical diagnosis of functional tic-

like behaviours: International consensus from experts

in tic disorders. European Journal of Neurology,

30(4):902–910.

pyradiomics community (2016). Radiomic features.

Rae, C. L., Critchley, H. D., and Seth, A. K. (2019). A

bayesian account of the sensory-motor interactions

underlying symptoms of tourette syndrome. Frontiers

in Psychiatry, 10.

Silva, J., Bispo, B., and Rodrigues, P. (2023). Structural

mri texture analysis for detecting alzheimer’s disease.

Journal of Medical and Biological Engineering.

Sun, J., Ovsjanikov, M., and Guibas, L. (2009). A con-

cise and provably informative multi-scale signature

based on heat diffusion. Computer Graphics Forum,

28(5):1383–1392.

Syed, A., Adam, R., Ren, T., Lu, J., Maldjian, T., and

Duong, T. Q. (2023). Machine learning with tex-

tural analysis of longitudinal multiparametric mri

and molecular subtypes accurately predicts pathologic

complete response in patients with invasive breast

cancer. PloS one, 18(1):e0280320.

Tourette.org, o. (2017). Iceberg illustration poster.

Yeh, F.-C. (2020). Shape analysis of the human association

pathways. NeuroImage, 223:117329.

Zhu, Y., Yang, D., Ji, W., Huang, T., Xue, L., Jiang,

X., Chen, L., and Wang, F. (2016). The relation-

ship between neurocircuitry dysfunctions and atten-

tion deﬁcit hyperactivity disorder: A review. BioMed

Research International, 2016:1–7.

Comparing 3D Shape and Texture Descriptors Towards Tourette’s Syndrome Prediction Using Pediatric Magnetic Resonance Imaging

481