Thyroid Ultrasound Images Classiﬁcation using the Shearlet Coefﬁcients

and the Generic Fourier Descriptor

Noura Aboudi

and Nawres Khlifa

Laboratoire de Biophysique et Technologies M

edicales, Universit

e de Tunis El Manar, Tunis, Tunisia

Keywords:

Thyroid Nodule, Feature Extraction, Shearlet Transform, Generic Fourier Descriptor, Feature Selection,

Random Forest.

Abstract:

To ameliorate the classiﬁcation accuracy of the thyroid ultrasound imaging computer-aided diagnosis (CAD)

system based on feature extraction, we used the Shearlet Transform (ST) to extract texture features, and the

Generic Fourier Descriptor (GFD) to extract shape feature descriptor based on contours information. The ST

supplies a rotation invariant descriptor at various scales. The GFD descriptor is autonomous, robust, and has

no redundant features. Then, we applied a feature selection method on the extracted shearlet descriptor to

build up the performance metrics. Finally, we used the objective metrics(sensitivity, speciﬁcity, and accuracy)

to validate the performance of the proposed method. Experimentally, we apply our novel methods on a public

dataset and we use the Support Vector Machine(SVM) and Random Forest (RF) as classiﬁer. The obtained

results prove the superiority of the proposed method.

1 INTRODUCTION

Thyroid nodule is an abnormal growth of cancerous

lumps in the thyroid gland, it is the accumulation of

malignant cells in thyroid gland tissues. It is one of

the most leading cause of cancer deaths. In 2017,

56,870 new patients in the United States have been

reported to have involved nearly thyroid cancer (Ab-

basian Ardakani et al., 2018). Generally, the most of

thyroid nodules diseases are benign.

Currently, thyroid ultrasound imaging has been

the most used tools for early thyroid nodules de-

tection and diagnosis. It is inexpensive, radiation-

free imaging tool, and provides the beneﬁts infor-

mation needed for medical diagnosis (Abbasian Ar-

dakani et al., 2018; Zhang and Lu, 2002). How-

ever, the diagnosis of thyroid ultrasound image de-

pends greatly on personal experience and skills. Thus,

many benign and malignant nodules have similar vi-

sual characteristics. Hence, experienced radiologists

have a high good diagnosis rate than beginner radi-

ologists. Thyroid ultrasound Computer Aided Diag-

nosis (CAD) system becoming progressively a cru-

cial tool, that assists to offers an objectivity evalu-

ation diagnostic and a better decision accuracy. In

general, the thyroid ultrasound CAD system is con-

stituted of four steps, containing image preprocess-

https://orcid.org/0000-0002-5101-853X

ing, image segmentation, feature extraction and se-

lection, and classiﬁcation. Feature extraction is one

of the crucial stages in thyroid ultrasound CAD. The

extracted features are usually classiﬁed into textural

and shape (morphological) features. Latterly, many

researches are focused on the feature extraction and

selection steps. Usually, the thyroid nodules classi-

ﬁcation problems depend on the extracted features.

Also, the textural features are commonly used in thy-

roid ultrasound CAD.

Usually, the textural features descriptors are cal-

culated using a diversity of statistical and structural

approaches, such as Grey Level Co-occurrence Ma-

trix (GLCM), autocorrelation-based approaches, Lo-

cal Binary Pattern, and auto-covariance coefﬁcients.

These methods can describe the statistical features of

grey level variation in a Region of Interest (ROI). The

popular advantage of these methods is they are easy to

implement, but the extracted textural features by these

methods are mostly from the special domain and ig-

nore the frequency features, which are very important

in the classiﬁcation step. Moreover, the multiscale

properties of an image not evaluated in these methods.

The Multiscale Geometric Analysis (MGA) grants

complete features analysis using different scales.

Recently, image analysis based on transform ap-

proaches has been widely used in the image feature

extraction. The transform approaches is the represen-

292

Aboudi, N. and Khlifa, N.

Thyroid Ultrasound Images Classiﬁcation using the Shearlet Coefﬁcients and the Generic Fourier Descriptor.

DOI: 10.5220/0008956902920298

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

292-298

ISBN: 978-989-758-402-2; ISSN: 2184-4321

tation of an image in the frequency and scale space;

when the features description and interpretation are

related on this special coordinate. The MGA is ap-

plied in different ﬁelds. One common ﬁeld is the tex-

ture feature extraction in thyroid ultrasound images.

The shearlet transform is a powerful spatial frequency

analysis method. Shearlet provides sufﬁcient tools

to exactly detect the orientations, the scales and the

positions of pixels (Easley et al., 2008a). Shearlet

transform has been fully utilized in image processing,

Edge and Ridge Detection and Analysis, image sep-

aration, and image denoising(?). Despite, a limited

research used the shearlet transform on the textural

features extraction.

Also, thyroid nodules analysis based on the shape

description is very important. Shape features can be

categorized into two principal groups: region-based

features and contour-based features. The contour-

based description method explores the boundary in-

formation and ignores the internal content of the

shape, so the versatility is not high. The region-

based description method uses the internal pixel in-

formation shape. Contour-based Descriptors can be

described by the fourier descriptor, the wavelet de-

scriptor, and the shapes signatures. In many exist-

ing shape feature descriptors, the Generic Fourier De-

scriptor (GFD) has several desirable features, such as

low computational complexity, sharpness to ﬁne de-

scription, which makes it a popular descriptor. The

GFD is one of the boundary feature extraction tech-

niques. GFD is obtained by applying the 2D Fourier

Transform in polar image and it extracts the spectral

features in radial and circular direction.

In this paper, we use the ST to extract textural fea-

ture and the GFD to explore shape descriptor features.

To achieve better classiﬁcation performance, we pro-

pose a hybrid approach combining ST and GFD for

ultrasound thyroid nodules classiﬁcation. The rest of

paper is organized as follows. In section 2, we in-

terpret the related work. The proposed method is de-

scribed in section 3. The section 4 represents the ob-

tained results and discussion. Finally, conclusion with

some feature works idea is given in section5.

2 RELATED WORKS

The calciﬁcation and detection of thyroid nodules in

ultrasound images was evolved in many studies. Dif-

ferent Computer-Aided Diagnosis was developed us-

ing a variety of features and classiﬁers for the clas-

siﬁcation of thyroid nodules. The most studies has

proven the potentiality and the importance of textural

and morphological features on the diagnostic of nod-

ules. Many CAD has developed for thyroid diseases

classiﬁcation. The neural networks has used by Ozy-

ilmaz et al.(Ozyilmaz and Yildirim, 2002) for the di-

agnosis of thyroid nodules, they are applied different

architectures on their database (13, 2017). The pro-

posed method attained 88.3% as maximal accuracy

value. Also, Keles et al.(Keles¸ and Keles¸, 2008), de-

veloped a CAD system based on neuro fuzzy clas-

siﬁcation testing on the similar dataset (13, 2017),

it attaining 95.33% accuracy value. Iakovidis et

al.(Iakovidis et al., 2010), proposed a method based

on textural and echogenicity features, focused on im-

age analysis. In this work, they used the fuzzy local

binary pattern to represent the texture feature. The

proposed method used 250 thyroid ultrasound im-

ages, achieving 97.5% as the best ROC AUC, utiliz-

ing polynomial kernel SVM as classiﬁer. Acharya et

Al.(Acharya et al., 2011), proposed a system for the

diagnosis and classiﬁcation of malignant thyroid nod-

ules using 20 contrast enhancement images(CEUS).

They used the Discrete Wavelet Transform (DWT)

and texture parameters for feature extraction. The

DWT detects the small variations in malignant and

benign nodules. The accuracy values achieve 98% us-

ing KNN classiﬁer. In another study(Acharya et al.,

2012), the same author combined a Fourier Descrip-

tor (FD), local binary patterns, fractal dimensions and

Law’s texture energy to detect features from 20 im-

ages. The highest accuracy value is of 100% us-

ing SVM and fuzzy classiﬁer. On other lately stud-

ies, in (Raghavendra et al., 2017) authors used the

Binary Stack Decomposition (BSD) algorithm and

Two-Threshold Binary Decomposition algorithm to

extract 120 features from 242 images. In this case,

a 97.52% accuracy value was attained using SVM

classiﬁer. The higher number of extracted features

decreases the performance and the exactitude of the

impact of these features. So, they have applied the

ﬁsher analysis (MFA) to select and reduce the fea-

tures sets. MFA based on the fuse of existing features

and the created features. Chi et al(Chi et al., 2017),

proposed a recently study published on 2017; they

used a deep learning features extraction method using

a GoogLeNet model. In this approach, 1024 features

was extracted and used to classify the nodules using

Random Forest classiﬁer. They used two dataset in

the evaluation of their method, the accuracy value at-

tained 98.29% of the ﬁrst dataset (357); and 96.34%

for the second dataset (164 images). The obtained re-

sults improves that deep learning offer a good results.

Thyroid Ultrasound Images Classiﬁcation using the Shearlet Coefﬁcients and the Generic Fourier Descriptor

293

Figure 1: Diagram of the proposed method.

3 MATERIALS AND METHODS

The proposed method is composed of three steps:

ROI selection, Shearlet Transform and GFD decom-

position feature extraction, feature selection, and clas-

siﬁcation based on SVM and RF. The proposed ap-

proach was presented in ﬁgure 1.

3.1 Data Collection

In our experiment, a total of 447 thyroid ultrasound

images with benign and malignant thyroid nodules

were used to evaluate the performance of the pro-

posed method. Thyroid ultrasound images used in this

work are acquired from the laboratory CIM @ LAB of

the National University of Colombia and the medical

diagnostic institute. Each image involves one or more

nodules (attached with XML ﬁle ready by expert and

contain annotation and patient’s information). The

ROI was selected using the position denoted in the

XML ﬁle. The original image had a resolution of

546*410. These extracted nodules are grouped into

two classes: benign and malign. In this work, out of

447 thyroid nodules, 372 nodules are malignant and

75 nodules are benign. An example of thyroid benign

and malign is presented in ﬁgure2 respectively.

Figure 2: Example of benign and malign thyroid nodules.

3.2 Feature Extraction

3.2.1 Shearlet-based Texture Feature

Descriptors

The textural feature are extracted using the Shearlet

Transfrom (ST) , it is brieﬂy deﬁned as follows:

Shearlet systems were ﬁrst introduced by K. Guo, G.

Kutyniok, D. Labate, W.-Q Lim and G. Weiss in (Guo

et al., 2006; Labate et al., 2005). ST is a multiscale

directional transform method which allows efﬁcient

encoding of anisotropic features, based on directional

ﬁlter followed by Laplacian pyramid (LP). Shearlet

Transform provides an effective tool for combining

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

294

the multi-scale and the invariance notation. This

multi-scale decomposition improves the robustness

multi-directional and multi-scale analysis and the

representation of the data image. It represent the

image in the frequency space where the texture

description is closely related to this coordinate. ST

used to identify directional features in images (Easley

et al., 2008b; Guo et al., 2006). It is generated by

applying a set of operators to a single function. The

most important advantage of the ST is not sensitive

with scales and orientations variations, and it is

more powerful in understanding the geometry of

images. Recently, ST constitute one of the most

successful methods for the efﬁcient representation

of multidimensional data and in the understanding

the geometry of images. Like that, it is only con-

nected with two parameters, the scaling parameter

a and the translation parameter t. For the input im-

age f , continuous Shearlet Transform is described by:

f −→ SH

f (a,s,t) = h f ,ψ

a,s,t

i (1)

Where:

ψ: is the generating function;

a : is the scaling parameter;

s ∈ R: shear parameter;

t ∈ R :translation parameter;

and ψ

a,s,t

:shearlet basis functions.

Important properties of shearlets are they are well lo-

calized, they follow parabolic scaling law, they have

high directional sensitivity, and they are optimally

sparse.

In the proposed method, ultrasound thyroid nod-

ule is decomposed by ST into three layers, therefore

the textural features descriptors were extracted from

these layers. The contrast, correlation, energy, homo-

geneity, entropy, skewness, variance, mean, standard-

deviation of each sub-bands are extracted from these

three layers for the horizontal and vertical cone and

are used as directional features. Shearlet can capture

directional features like orientations in images, which

are in fact one of the most discriminating features. By

practicing the ST to an image, a number of decompo-

sition levels and directional subbands are produced.

3.2.2 Fourier Feature Descriptors

In this study, we have utilized the generic Fourier

descriptor (GFD) for efﬁcient shape representation.

Most of the actual shape descriptors are non-robust.

GFD is proposed to crucify the disadvantages of

the current shape representation methods that inde-

pendent, easy to implement, less sensitive to noise,

and robust. GFD introduced by Zhang(Zhang and

Lu, 2002). It is a contour-based shape descriptor for

image classiﬁcation. To obtain invariance to rotation,

the image is ﬁrst converted to polar coordinates then

we use a 2D Fourier transform on a polar image.

GFD uses the modiﬁed polar Fourier transformation

(MPFT) of a region shape to the polar coordinate

system. So, the coordinates of all pixels of the

initial images are converted into polar coordinates.

It detects spectral feature in both radial and circular

direction. The determination of the number T and

R for the description of the forms is physically

feasible, because the shape characteristics are usually

extracted by the low frequencies. Finally, the GFD

has the following expression:

For an shape image

I = { f (x,y);0 ≤ x < M,0 ≤ N} (2)

p f (ρ, ψ) =

∑

r=0

∑

i=0

f (r,θ

)

j2π(

ρ+

2πi

ψ)

(3)

Where:

T : is the angular resolution;

R : is the radial resolution;

The Fourier coefﬁcients obtained are translation in-

variant. So to realize scaling and rotation invariance,

the following normalization is calculated:

GFD = {|

p f (0, 0)

aire

p f (0, 1)

p f (0, 0)

,..., |

p f (0, n)

p f (0, 0)

,..., |

p f (m, n)

p f (0, 0)

}

(4)

Where:

area : symbolize the area of the border circle in which

the polar image exists;

andn= max (angular frequencies);

m = max (radial frequencies).

In this paper, we have used the Generic Fourier

Fescriptor (GFD) as shape descriptor. For efﬁcient

shape description, only a small number of GFD fea-

tures are selected for shape representation. In our

implementation, 36 GFD features reﬂecting 4 radial

frequencies and 9 angular frequencies are selected to

index the shape. The selected GFD features form a

feature vector which is used for indexing the shape.

Therefore, the online matching is efﬁcient and sim-

ple. The extracted features with GFD are no redun-

dant and it grants a multi-resolution feature analy-

sis in radial and angular directions. The GFD fea-

ture vector is introduced by the following expression:

GFD(0, 0),GFD(0, 1), . . . ,GFD(0, n), . . . ,GFD(m,

0). . . ,GFD(m, n).

3.3 Feature Selection

Feature selection is a dimensionality reduction

method which aims to select a subset of relevant and

Thyroid Ultrasound Images Classiﬁcation using the Shearlet Coefﬁcients and the Generic Fourier Descriptor

295

informative features from the initial features set by

eliminating irrelevant and redundant features. It aug-

ments the classiﬁcation performance, the efﬁciency in

learning stage, and reducing the computation cost and

the complexity. The total number of extracted texture

feature was 207. Some of these are irrelevant and not

signiﬁcant in the differentiation between benign and

malignant nodules and not appropriate for classiﬁca-

tion. The feature selection method utilized should be

able to choose a subset of relevant and most repre-

sentative features. In this study, we have applied the

Reliable Attribute Selection Based on Random Forest

(RASER)(Noura et al., 2016) to eliminate the non-

informative features. The features results obtained af-

ter applying this feature selection method was used

in the classiﬁcation, and for construct SVM and RF

models.

3.4 Classiﬁcation Algorithm

In order to evaluate the performance of the proposed

method, the extracted features were fed to the classi-

ﬁer for discriminating the malignant from the benign

nodules. Machine Vector Support (SVM) and Ran-

dom Forest (RF) were applied to measure the perfor-

mance of these features. Concise descriptions of these

two classiﬁers are given below.

3.4.1 Support Vector Machine

Support Vector Machine (SVM) is a supervised learn-

ing algorithm, SVM belongs to the class of linear

classiﬁers (that use a linear separation of data). It

is known for their strong theoretical guarantees, their

great ﬂexibility and their ease of use even without

much knowledge of data. SVM is intended to sep-

arate data into classes using a boundary, so that the

distance between the different groups of data and the

boundary between them is maximum. SVM is based

on the generation of hyperplanes to discriminate fea-

tures categorizing to two different classes. In this

study, a weighted SVM algorithm is used to equili-

brate imbalance classes. Weighted SVM (W-SVM)

solves the problem of having two classes with unequal

training data. W-SVM sets the penalty parameter C

in proportion to the size of the class. With regard to

RF, the principal advantage of SVM is its simpler ge-

ometric interpretation and lower computational cost.

The main advantages of SVM are its simpler inter-

pretation and computation cost compared to Random

Forest (RF).

3.4.2 Random Forest

Random Forest (RF) is a one of the ensemble meth-

ods of classiﬁcation. The ensemble methods type is

based on vote to predict the ﬁnal decision. RF is con-

structs of a large number of decision trees based on

averaging random selection of variables. RF is based

on the idea of bagging and Random subspaces in the

construction of decision trees. The randomness no-

tion is in the subsampling of the training data and in

the selection of the node tests, each tree is build us-

ing different subset. RF uses the majority votes in the

classiﬁcation case in the terminal leaf nodes. More

than, RF has the ability to measure the importance of

used the features.

3.4.3 Validation of Classiﬁer

We have used the group 10 fold cross-validation on

the evaluation of the proposed method. K-fold cross

validation based on the random split of the dataset

into k equal samples, and it guarantee that the same

set of data not be selected in both testing and train-

ing sets. Between the k samples, one sample is used

as test dataset and the remaining k-1 sets are used as

training dataset. The accuracy, sensitivity, speciﬁcity

were chosen as critical measure performance of the

proposed method for both SVM and random forest

classiﬁers. Their deﬁnitions are as follows:

Sensitivity = T P/(T P + FN) (5)

Speci f icity = T N/(T N + FP) (6)

Accuracy =

T P + T N

T P + T N + FP + FN

(7)

where : T P represent the correct classiﬁcation rate of

malignant instances;

FN is the misclassiﬁcation rate of malignant in-

stances;

T N is the correct benign instances; and

FP indicates the misclassiﬁcation rate of benign in-

stances.

4 RESULTS AND DISCUSSION

In this paper, the thyroid nodules are classiﬁed into

two classes: benign and malignant. The ROIs are

extracted initiallly from the ultrasound images and

then subjected to the shearlet transform and generic

fourier descriptor. All nodules are decomposed with

shearlet transform into three layers. The 9 textu-

ral features(contrast, correlation, energy, homogene-

ity, entropy, skewness, variance, mean,and standard-

deviation) are extracted from each sub bands. In total,

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

296

Table 1: Classiﬁcation performance for Shearlet feature using SVM and RF(unit: %).

Index Without feature selection With feature selection

SVM RF SVM RF

Accuracy 93.3 94.65 94.59 96.64

Sensitivity 89.86 92.95 91.78 93.15

Speciﬁcity 91.6 95.79 93.11 95.98

Table 2: Classiﬁcation performance for different features using SVM and RF (unit:%).

Index Shearlet GFD Fusion feature

SVM RF SVM RF SVM RF

Accuracy 94.59 96.64 97.74 95.9 96.55 98.51

Sensitivity 91.78 93.15 97.58 97.4 95.87 98.17

Speciﬁcity 93.11 95.98 98.11 89.2 97.11 96.08

207 features were extracted from the all sub bands.

Further, these features were subjected to Reliable At-

tribute Selection Based on Random Forest (RASER)

dimensionality reduction method to eliminite redun-

dant and irrelivant features. Finally, relevant features

were fed to SVM and RF classiﬁer to test the pro-

posed method based on 10 cross validation. The table

1 represents a comparison of the classiﬁcation per-

formance using SVM and RF classiﬁer. The feature

selection method provides also a good classiﬁcation

performance compared to the classiﬁcation without

feature reduction method.

It can be note that best accuracy, sensitivity and

speciﬁcity values are 94.59%, 91.78%, and 93.11%

for SVM and 96.64%, 93.15%, and 95.98% for RF

after apply the feature selection. In rest of study,

only the relevant shearlet coefﬁcient are used. Unlike

shearlet transform, GFD has no redundant features.

Table 2 represents the obtained evaluation metrics

value between different feature type using SVM

and RF classiﬁers, respectively. GFD achieves

better classiﬁcation compared to the Shearlet-based

method. The accuracy value achieve 97.74%,96.64 %

respectively for GFD using SVM and RF classiﬁer,

and 94.59%, 95.9% for the shearlet coefﬁcient using

SVM and RF.

Therefore, the quantitative results of classiﬁcation

accuracy, sensitivity, and speciﬁcity for different

combination of texture features and classiﬁers are

shown in table 2. They show that the contribution of

the new descriptors improves the overall accuracy,

sensitivity, and speciﬁty. The classiﬁcation accuracy

of the fusion features is 96.55% and 98.51% using

SVM and RF respectively, which are much higher

than those of other methods.

Recently, several researchers have encouraged to

propose new efﬁcient method to diagnose thyroid

cancer using ultrasound images. Table 3 summarizes

the achieved results on thyroid nodule classiﬁcation,

we introduce the obtained performances by present-

ing the accuracy value. The compared methods are:

Conic Section Function Neural Network (Ozyilmaz

and Yildirim, 2002); neuro fuzzy Classiﬁcation(Keles¸

and Keles¸, 2008); fuzzy local binary pattern(Iakovidis

et al., 2010); Discrete Wavelet Transform (DWT)

and texture parameters(Acharya et al., 2011) that

is based in the texture feature and the wavelet

transform; Two-Threshold Binary Decomposition

algorithm(Acharya et al., 2012); and deep learning

features extraction(Chi et al., 2017). Really, we

produced the accuracy value for the proposed method

as well as for six relevant thyroid nodules classiﬁ-

cation methods from the state of the art. It can be

clearly seen that the proposed method reaches a good

performances and still better to other methods.

In this paper, we have proposed an efﬁcient

method for classiﬁcation of thyroid nodules using the

shearet transform, generic fourier descriptor and in-

variant texture features. Shearleat transform has some

important properties, like multiresolution, multidirec-

tion and multiscale, which approve the uniqueness

above different levels. At present, textural feature is

commonly used in CAD system to classify the ultra-

sound thyroid nodules. The combination of statistical

and transform based features improve the classiﬁca-

tion accuracy for thyroid nodules classiﬁcation.

5 CONCLUSION

In conclusion, we proposed a new feature extrac-

tion method based on shearlet transform and generic

fourier descriptor for characterizing thyroid nodules

in ultrasound image. The comparative experiment re-

sults indicated that the combination of both shearlet-

based texture and fourier based edge features have

the best classiﬁcation performance. The proposed

method was tested on public thyroid database re-

Thyroid Ultrasound Images Classiﬁcation using the Shearlet Coefﬁcients and the Generic Fourier Descriptor

297

Table 3: Evaluation of the proposed method comparatively to the other methods (unit: %).

Methods Accuracy

Ozyilmaz et al(2002) Conic Section Function Neural

Network

88.3

Keles et al(2008) neuro fuzzy Classiﬁcation 95.33

Iakovidis et al(2010) fuzzy local binary pattern 97.5

Acharya et Al(2011) Discrete Wavelet Transform

(DWT) and texture parameters

Acharya et Al(2017) Two-Threshold Binary Decom-

position algorithm

97.52

Chi et al(2017) deep learning features extrac-

tion

98.29

Proposed methods Shearlet Transform and

Generic Fourier Descriptor

98.51

quired from the laboratory CIM @ LAB. The classiﬁ-

cation performance of textural feature has also been

optimized by the RASER dimensionality reduction

method. A comparative study shows that the metrics

performance is better with the application of feature

selection step. In addition, GFD was used to extract

the boundary information. Finally, the combination of

texture feature obtained with shearlet decomposition

and boundary information acquired using GFD give

the highest classiﬁcation performance.

Thus, in our future work, we want to propose an

automated CAD system for detection and classiﬁca-

tion of thyroid nodules.We aim also to study and pro-

pose other feature selection methods.

REFERENCES

(2017). Uci repository of machine learning databases, uni-

versity of columbia at irvine.

Abbasian Ardakani, A., Reiazi, R., and Mohammadi, A.

(2018). A clinical decision support system using ultra-

sound textures and radiologic features to distinguish

metastasis from tumor-free cervical lymph nodes in

patients with papillary thyroid carcinoma. Journal of

Ultrasound in Medicine, 37(11):2527–2535.

Acharya, U. R., Faust, O., Sree, S. V., Molinari, F., Gar-

beroglio, R., and Suri, J. (2011). Cost-effective and

non-invasive automated benign & malignant thyroid

lesion classiﬁcation in 3d contrast-enhanced ultra-

sound using combination of wavelets and textures: a

class of thyroscan

algorithms. Technology in can-

cer research & treatment, 10(4):371–380.

Acharya, U. R., Sree, S. V., Krishnan, M. M. R., Moli-

nari, F., Garberoglio, R., and Suri, J. S. (2012). Non-

invasive automated 3d thyroid lesion classiﬁcation in

ultrasound: a class of thyroscan

systems. Ultrason-

ics, 52(4):508–520.

Chi, J., Walia, E., Babyn, P., Wang, J., Groot, G., and

Eramian, M. (2017). Thyroid nodule classiﬁcation

in ultrasound images by ﬁne-tuning deep convolu-

tional neural network. Journal of digital imaging,

30(4):477–486.

Easley, G., Labate, D., and Lim, W.-Q. (2008a). Sparse

directional image representations using the discrete

shearlet transform. Applied and Computational Har-

monic Analysis, 25(1):25–46.

Easley, G., Labate, D., and Lim, W.-Q. (2008b). Sparse

directional image representations using the discrete

shearlet transform. Applied and Computational Har-

monic Analysis, 25(1):25–46.

Guo, K., Kutyniok, G., and Labate, D. (2006). Sparse multi-

dimensional representations using anisotropic dilation

and shear operators.

Iakovidis, D. K., Keramidas, E. G., and Maroulis, D.

(2010). Fusion of fuzzy statistical distributions for

classiﬁcation of thyroid ultrasound patterns. Artiﬁcial

Intelligence in Medicine, 50(1):33–41.

Keles¸, A. and Keles¸, A. (2008). Estdd: Expert system for

thyroid diseases diagnosis. Expert Systems with Ap-

plications, 34(1):242–246.

Labate, D., Lim, W.-Q., Kutyniok, G., and Weiss, G.

(2005). Sparse multidimensional representation using

shearlets. In Wavelets XI, volume 5914, page 59140U.

International Society for Optics and Photonics.

Noura, A., Shili, H., and Romdhane, L. B. (2016). Reliable

attribute selection based on random forest (raser). In

International Conference on Intelligent Systems De-

sign and Applications, pages 11–24. Springer.

Ozyilmaz, L. and Yildirim, T. (2002). Diagnosis of thy-

roid disease using artiﬁcial neural network methods.

In Proceedings of the 9th International Conference on

Neural Information Processing, 2002. ICONIP’02.,

volume 4, pages 2033–2036. IEEE.

Raghavendra, U., Acharya, U. R., Gudigar, A., Tan, J. H.,

Fujita, H., Hagiwara, Y., Molinari, F., Kongmebhol,

P., and Ng, K. H. (2017). Fusion of spatial gray level

dependency and fractal texture features for the charac-

terization of thyroid lesions. Ultrasonics, 77:110–120.

Zhang, D. and Lu, G. (2002). Generic fourier descriptor for

shape-based image retrieval. In Proceedings. IEEE In-

ternational Conference on Multimedia and Expo, vol-

ume 1, pages 425–428. IEEE.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

298