Original

Implementation of Combinational Deep MSc, Department of

1

Medical Physics and

Learning Algorithm for Non-alcoholic Biomedical Engineering,
School of Medicine, Sha-
Fatty Liver Classification in Ultrasound hid Beheshti University
of Medical Sciences,
Images Tehran, Iran
2
PhD, Department of
Medical Physics and
Biomedical Engineering
Zamanian H.1 , Mostaar A.2,3* , Azadeh P.4, Ahmadi M.2 and, School of Medicine,
Shahid Beheshti Univer-
sity of Medical Sciences,
ABSTRACT Tehran, Iran
3
PhD, Radiation Biology
Background: Nowadays, fatty liver is one of the commonly occurred diseases for Research Center, Iran
the liver which can be observed generally in obese patients. Final results from a vari- University of Medical Sci-
ety of exams and imaging methods can help to identify and evaluate people affected ences, Tehran, Iran
by this condition. 4
MD, Department of Ra-
diation Oncology, School
Objective: The aim of this study is to present a combined algorithm based on of Medicine, Shahid
neural networks for the classification of ultrasound ‎images from fatty liver affected Beheshti University of
patients. Medical Sciences, Teh-
Material and Methods: In experimental research can be categorized as a ran, Iran
diagnostic study which focuses on classification of the acquired ultrasonography
images for 55 patients with fatty liver. We implemented pre-trained convolutional
neural networks of Inception-ResNetv2, GoogleNet, AlexNet, and ResNet101 to
extract features from the images and after combining these resulted features, we pro-
vided support vector machine (SVM) algorithm to classify the liver images. Then the
results are compared with the ones in implementing the algorithms independently.
Results: The area under the receiver operating characteristic curve (AUC) for the
introduced combined network resulted in 0.9999, which is a better result compared to
any of the other introduced algorithms. The resulted accuracy for the proposed net-
work also caused 0.9864, which seems acceptable accuracy for clinical application.
Conclusion: The proposed network can be used with high accuracy to classify
ultrasound images of the liver to normal or fatty. The presented approach besides the
high AUC in comparison with other methods have the independence of the method
from the ‎user or expert interference.
Citation: Zamanian H, Mostaar A, Azadeh P, Ahmadi M. Implementation of Combinational Deep Learning Algorithm for Non-alcoholic Fatty
Liver Classification in Ultrasound Images. J Biomed Phys Eng. 2021;11(1):73-84. doi: 10.31661/jbpe.v0i0.2009-1180.

Keywords
Fatty Liver; Ultrasonography; Deep Learning; Transfer Learning; Support
Vector Machine; Receiver Operating Characteristic Curve
*Corresponding author:
A. Mostaar

T
Department of Medical
Introduction Physics and Biomedi-
he liver is one of the crucial body organs that as a part of the di- cal Engineering, School
gestive system has an irrefutable rule in the chain of digesting the of Medicine, Shahid
Beheshti University of
dietary elements. This vital organ contains commonly hepatocyte Medical Sciences, Teh-
which controls and regulates a high volume of biochemical activities, ran, Iran
including synthesis and breaking complex and tiny molecules [1]. In E-mail: mostaar@sbmu.
ac.ir
relative methods for medical imaging, experts usually pay attention to
Received: 12 September 2020
find, assign the conditions, and monitor patients based on experience Accepted: 23 October 2020

J Biomed Phys Eng 2021; 11(1) 73

Zamanian H. et al
and profession. This way is usually personal proach to evaluate liver textures by using ul-
and inaccurate [2]. Thus, implementing arti- trasound B-mode images [11]. They extracted
ficial intelligence (AI) approaches have been some of the image features and then used the
common to increment the accuracy of diag- artificial neural network (ANN), SVM, and
noses [3]. The AI and machine learning algo- k-nearest neighbor (kNN) methods for the
rithms made efforts that besides concentrating classification. Meanwhile, their algorithm re-
on different processes such as preprocessing, quires assigning a region of interest (ROI) by
image segmentation based on color or texture specialist, thus it needs to access a specialist
specification, have activities on extracting fea- and also ‎based on different comments, defin-
tures and image classification. In this way, a ing a comprehensive gold standard region is
special type of neural networks, called convo- difficult. In 2014, Gao et al. used the GLCM
lutional neural network (CNN), shows a com- algorithm to identify 22 features with the most
prehensive success in the analyses and image varieties from the acquired ultrasound images
recognition application. A variety of applica- of liver texture with the goal of classification
tions has been considered for these elements, [12]. This rate of variation can be explained
including structural models, pattern recogni- by pathological specification and the outward
tion, and predictive procedures [4-7]. Basi- form of the texture. But again, in this algo-
cally, the learning algorithms operate based rithm, the subject of convergence and calcula-
on existing characteristics for the system. On tion of learning error can be questionable. In
the other words, their operating goal is to learn 2016, Achyara and his group suggest a strat-
the effective features of the existed data for the egy to distinguish different classes by curvelet
activities that this learning procedure is called transform algorithm on the ultrasound B-mode
deep learning [8]. In 2005, Cao et al. extract- images [13]. This group used entropy features
ed the features by the use of common fractal from CT coefficients and its effective ones
dimension specifications and texture’s edge were selected by locally sensitive discrimi-
co-occurrence matrix from ultrasound images nant analysis (LSDA). After that, this feature
and then used linear classification algorithms was graded by F-value, and their classification
fisher and Support Vector Machine (SVM) to was done by different networks. Although the
categorize liver texture to normal or abnormal importance of features was considered, there
[9]. Due to the desired performance based on was no comprehensive evaluation of all speci-
the condition of acquired information, this al- fications of the images and there may be more
gorithm couldn’t extract generally all features valuable features for using in the classifica-
and the classification is not based on whole tion. In 2017, Kuppili et al. presented a novel
dominant specifications of the images. In algorithm based on a fast Extreme learning
2012, Acharya et al. proposed a combinational machine (ELM) to classify the rate of steato-
algorithm to extract features of ultrasound im- sis in ultrasound images of the liver [14]. The
ages of the liver in which, the wavelet trans- resulted features with 46 other specifications
form method and higher-order spectra of the of the images were used for classification by
images were used to extract the effective fea- the ELM method. In this algorithm, besides
tures and the classification was used by fuzzy an appropriate accuracy, its complexity and
and SVM classifiers [10]. But this algorithm necessity to have a gold standard for evalua-
depends on the number of applicative features tion make it difficult to be applicable. In 2017,
and the greater number of input features and Hassan et al. offered a chained separated au-
also based on its ‎classification, the value of fi- to-encoder (SAE) algorithm to extract image
nal accuracy will vary. In 2012, Andrade et al. features for focal diseases [15]. The region of
proposed a semi-automatic classification ap- interest was segmented by the level-set algo-

74 J Biomed Phys Eng 2021; 11(1)

 Combinational Deep Learning Algorithm for NAFLD Classification
rithm and k-means clustering methods, and and SVM method to classify them to assess
the SAE algorithm extracts high-level features the fat steatosis in non-alcoholic fatty liver
of these areas. Then the SoftMax layer is re- patients [19]. They used acquired ultrasound
sponsible for classifying them for different images from 55 patients admitted for bariat-
maladies. The result of this algorithm is evalu- ric surgery (laparoscopic sleeve gastrectomy).
ated appropriately in comparison with other Byra et al. compared their proposed algorithm
common methods, but its dependency on the with two previously common methods [20],
segmentation method and high mathematical but this was done based on just one transfer
complexities for statistical calculations make learning algorithm. Besides, in their proposed
the use of this method questionable. In 2017, algorithm the extracted features provided by
Liu et al. sculptured an algorithm in which its HI and GLCM methods were ‎based on appro-
feature extraction and classification are pro- priate specialist and experience in imaging,‎
vided from a selected collection of regions type of US system, configuring of US imaging
in a segmented capsule from the liver images system, assigning the region of interest, and
[16]. Geometrical specifications of the as- so on which may result in completely different
signed capsuled area of the liver can be used outcomes for different execution and special-
as an introduction to diagnose the image class, ists. Various attempts summarized from their
but determining these features needs a previ- efforts are to estimate fatty liver disease ac-
ous experience and completes familiarity with cording to a learning algorithm and in a deter-
the condition. Apart from that, the considered mined ROI [21-24].
size of the window has influenced on the clas- This article aims to implement a combina-
sification result. In 2017, Bharath and Rajal- tional deep learning algorithm for classify-
akshmi used the invariant scattering convo- ing the level of fat steatosis based on transfer
lutional network to extract the features of the learning and for the B-mode ultrasound ac-
ultrasound images of the liver and assigned quired images of liver texture. The ultrasound
its fatty level, and they accordingly imple- images applied for this work are from some se-
mented SVM classification to categorize the verely obese patients, collected before bariat-
results [17]. The operation of this algorithm is ric surgery. In this work, we used pre-trained
based on the frequency of the variation in its convolutional neural networks to extract the
data that if they are, the differentiation power features of the images. Then they are used for
will be more, which it is not always achiev- image classification by the SVM method by
able. In 2018, Biswas et al. propounded a net- combining these features. In the following, the
work based on CNN structure with different performance of the proposed algorithm in ac-
layers of convolution, pooling, and inception curacy, sensitivity, and specificity is compared
for characterizing liver texture and assign- with the result of each network, solely. The
ing the risk level for getting infected by fatty total structure of this paper is divided as fol-
liver hepatitis under symtosis class [18]. This lows; firstly, we describe the group of patients
algorithm extracts its features of ROI remov- used for imaging and the specification of ap-
ing background information and provides this plicable data. Secondly, the proposed combi-
region according to a gold truth. This method national deep learning algorithm is introduced
is just applicable to organs which their gold for assessing non-alcoholic fatty liver. In this
truth exists. This obviously shows demand part, the procedure of using a transfer learn-
for an expert like previous methods. In 2018, ing algorithm is presented for extracting the
Byra et al. presented a learning algorithm by features of the images by using CNN. Then
using of pre-trained Inception-ResNetV2 neu- the provided results of the previous section
ral network to extract features of the images are employed to classify and assess steatosis

J Biomed Phys Eng 2021; 11(1) 75

Zamanian H. et al
for fat in the liver hepatitis. Finally, the results parameter, the fatty liver is defined as having
are presented and we discuss this algorithm’s the level of steatosis more than 5% [25]. Thus,
advantages and disadvantages. The aim of the described patients are categorized into 2
this study is to present a combined algorithm classes of susceptible to fatty liver (steatosis
based on neural networks for the classification more than 5%) and normal people (less than
of ultrasound ‎images from fatty liver affected 5%). Figure 1 shows an image of normal liver
patients. and an image of susceptible to the fatty liver
as a sample. Due to motion and relative posi-
Material and Methods tion of the liver and kidney, corresponding to
In this experimental study, we used of ac- one heartbeat, a sequence of images, including
quired ultrasound images from 55 patients ad- 10 slightly different images are acquired and
mitted for bariatric surgery (mean age 40, 20% stored for each patient. Finally, 550 B-mode
male, mean Body Max Index (BMI) 45.9) ultrasound images construct the related datas-
[19]. They have been admitted for gastrec- et used for the following investigation. Figure
tomy surgery by laparoscopy. This imaging 2 shows the distribution histogram for the ac-
has been taken by Internal Medicine, Hyper- quired information of the volunteers. Accord-
tension and Vascular Diseases, Medical Uni- ing to this histogram, the data acquisition pro-
versity of Warsaw, Poland, during the cardiac cedure has been done for 38 patients suffering
echocardiographic evaluation, 1-2 days before from fatty liver and 17 normal people.
their surgery. Each patient sustains a wedge The acquired images have been presented in
liver biopsy during the surgery as a part of the size of 434 × 636 pixels (for each pixel with
protocol assigned at the Department of Gen- the size of 0.373×0.373 mm2), for different pa-
eral, Transplant, and liver surgery, Medical tients. Based on database information, it can
University of Warsaw, Poland, and therefore, be considered with two problems; firstly, for
the level of steatosis for each patient has been using these images as the input data for pre-
determined based on this sampling. Pathologic trained networks, the size of whole images
evaluation of these samples has been done by must be configured based on the condition of
a pathologist, who follows the protocols and the network’s input layer. Besides, the images
comments of the clinical research network for volunteer patients are in unequal numbers
[25]. According to this evaluation, the level of for different classes. To solve the first, the im-
steatosis is defined by the percentage of he- ages are resized on the adaptive dimension,
patocytes with fatty infiltration. Based on this described for the relative network. Therefore,

Figure 1: Ultrasound image samples for patients: a) normal liver (Index 3%), b) susceptible to
fatty liver (Index 20%).

76 J Biomed Phys Eng 2021; 11(1)

 Combinational Deep Learning Algorithm for NAFLD Classification
fashion is used for plenty of deep learning
networks as a fast and reliable method [26].
In this paper, we used rotation technique to
assimilate the number of images in the class
of normal people so that each image just has
2 degrees counter-clockwise rotation. By this
method, the number of images will be compa-
rable for two classes.
In the next step, we used some different pre-
trained networks to extract images features.
These networks are Inception-ResNetV2,
GoogleNet, AlexNet, and ResNet101. The In-
ception-ResNetv2 neural network pre-trained
Figure 2: Distribution Histogram of the level by the ImageNet dataset has 825 total layers
of steatosis versus the population of tested and it provides the property of decreasing
patients. computational volume. In this network, it is
considered that many activities in deep neu-
ral networks are redundant because there is a
the processing of images collection for pre- correlation between them in neighbors. Thus
trained networks will be easily possible. Sec- some of the similar activities in feature extrac-
ondly, we use image augmentation methods. tion can be decreased or removed by dropout
In these methods, a new information series layers. AlexNet network composes of 5 con-
is generated from the previous ones, with- volutional neural networks that follow with
out adding any new data. In other words, no 3 fully connected layers. This network uses
new images with different features are added of ReLU function instead of tangent or sig-
to the collection, but the number of existing moid ones in neural network structures, just
elements and the quantity weight of informa- like other CNN. This privilege provides more
tion is balanced in different classes with just speed on training, in comparison with two
changing the conjunction of existed features previous functions. After extracting different
to equalize the possible effect of all classes for features of the images by different mentioned
selection. There are different proposed algo- networks, they are aggregated to each other
rithms for this such as traditional transforma- and result in a comprehensive information
tion, GAN network, texture transfer, random bank of the possible features. Each network
erasing technique, random noise augmentation results in 1000 features at final layers that may
technique and so forth. One of the modest ex- be the same with one of the other networks.
isting methods is the traditional transfer of the This will not cause any disruption to the final
images. In this common method, a combina- classification since the dedicated weights to
tion of image transformation and color modi- the specification of each image are normalized
fication in the existed images is done and the in the provided statistical calculation space for
results are stored as new ones. Some of these classifying. In other words, the weight effect
transformations are rotation, mirror, rescaling of the repetitive features is distributed in all
(zooming), and cropping. Changing the color images equally.
of images is obtained in different algorithms After extracting existing features using these
such as histogram equalization, contrast incre- networks, the SVM algorithm is employed for
ment or lightning of images, brightness equal- classification. In the SVM method, the dis-
ization of images, blurring and so on. This tance of the features is measured toward a lin-

J Biomed Phys Eng 2021; 11(1) 77

Zamanian H. et al

ear or nonlinear kernel function and the com- TN

Specificity = (2)
parison criterion for estimating data in each TN + FP
class has been determined as their distance to Accuracy: The accuracy indicator is de-
the support vector. This vector is rated as the scribed by all the number of all elements clas-
closest distance founded from each class to- sified correctly to all existing elements for
ward optimal kernel function. Figure 3 shows classification and it can be calculated as fol-
the mounting structure of information versus a lowing:
linear kernel function. TP +TN
The evaluation standards for accuracy of Accuracy = (3)
TP +TN + FP + FN
output in applicable networks are different.
The main criteria are accuracy, sensitivity, and The criteria of accuracy, sensitivity, and
specificity. specificity are calculated for the presented al-
Sensitivity: Sensitivity is defined as the gorithm. Generally, the confusion matrix can
percentage of elements correctly classified in be an appropriate summation to express the
class 1 (True Positive (TP)) to all elements la- condition of these parameters for test data to
beled in class 1 (True Positive (TP) + False compare their performance. In this work, the
Negative (FN)) and it is calculated as follow- evaluation and correctness of the implement-
ing: ed algorithm are guaranteed by 10-fold cross
TP validation and the rate of errors between their
Sensitivity = (1) results is presented as a canon. In each col-
TP + FN
lection, the algorithm considers 75% of the
Specificity: Specificity can also be described presented information for the training proce-
by the percentage of elements classified cor- dure and uses the rest for testing the result.
rectly in class 2 (True Negative (TN)), in com- After training, the posteriori probabilities are
parison with all existed elements in class 2 calculated for each randomly selected test im-
(True Negative (TN) + False Positive (FP)), ages and they are exploited for extracting the
and it can be presented as following: receiver operating characteristic (ROC) curve.
The posteriori probabilities are a statistical
quantity after receiving data. In other words,
the posteriori probability is the conditional
probability of a quantity provided that it has
been seen that estimated data. Simply speak-
ing, this parameter provides the possibility of
substituting data in a certain class. Moreover,
the Area under the ROC Curve (AUC) is uti-
lized for evaluating the performance of the
classification procedure.

Results
In summarize, the overall procedure of this
proposed algorithm is shown in Figure 4.
The performance of the proposed network
was calculated by the SVM classification meth-
Figure 3: The structure of information in the od and its result has shown in Figure 5. These
Support vector machine (SVM) classification results are presented for comparison with the
structure. outcomes of employing pre-trained networks

78 J Biomed Phys Eng 2021; 11(1)

 Combinational Deep Learning Algorithm for NAFLD Classification

Figure 4: The overall block diagram of the implemented algorithm.

Figure 5: The receiver operating characteristic (ROC) curves for different applicable networks
in classification; proposed combinational neural network, Inception-ResNetV2, GoogleNet,
AlexNet, ResNet101.

mentioned above solely and with the SoftMax and GoogleNet neural networks are approxi-
classification layer. The whole applicable net- mately the same, meaning equal by 0.996. In
works have shown appropriate performances. detailed view, Figure 6 provides configura-
According to Figure 5, the highest AUC, equal tion matrix for different presented algorithms
to 1, was obtained corresponding to classifi- on the target data. The overall summary of
cation with ResNet101 pre-trained neural net- results for the performance of different men-
work. After that, with very little difference, the tioned networks is presented in Table 1, for
proposed algorithm has shown the maximum comparing with the proposed algorithm. The
AUC, meaning 0.9999. The lowest AUC was accuracy, sensitivity, and specificity values for
obtained using the Inception-ResNetV2 net- the applicable networks are assigned and pre-
work that was 0.9757. The results for AlexNet sented according to the ROC curve. Table 2

J Biomed Phys Eng 2021; 11(1) 79

Zamanian H. et al

Figure 6: Comparison of confusion matrices for: a) the proposed algorithm, b) Inception-

ResNetV2, c) GoogleNet, d) AlexNet, e) ResNet101 networks.

Table 1: Summarize performance results for different Networks.

Type of Network Specificity (%) Sensitivity (%) Accuracy (%) AUC K-fold loss
Inception-ResNetV2 63.2 100 81.08 0.9757 4.22e-2
GoogleNet 89.5 100 94.6 0.9960 4.05e-2
AlexNet 100 98.6 99.32 0.9963 3.54e-2
ResNet101 100 98.6 99.32 0.9998 3.54e-2
Proposed Algorithm 100 97.20 98.64 0.9999 3.40e-3
AUC: Area Under Curve

displays a comparison of different algorithms tients, which are affected by excessive obesity
with the proposed one to elucidate the perfor- and also are a candidate for bariatric surgery. In
mance of the method clearly. following the procedure of this paper, at first,
the acquired images are modified on their size
Discussion and sample numbers of each class, in prepro-
As mentioned above, ultrasound imaging is cessing step, and after that, their features are
considered as one of the common methods for extracted by Inception-ResNetV2, GoogleNet,
medical diagnostic procedures. In this paper, AlexNet, and ResNet101 pre-trained neural
we tried to find a classification function with networks. According to different behavioral
appropriate accuracy, based on these images implicit of applicable networks, the resulted
and the assigned labels considered by biopsy features are different and convey various as-
samplings. In this way, we used the ultrasound pects of the images. Therefore, in the next
images of liver textures for 55 different pa- step, these features are merged together and

80 J Biomed Phys Eng 2021; 11(1)

 Combinational Deep Learning Algorithm for NAFLD Classification
then they are used for classification by SVM was used for classification, the results show
algorithm in normal and fatty liver classes. these data can display the capability of transfer
The high level of the resulted performances for learning networks very well. The AUC index
the proposed algorithm and different networks was obtained 0.9999 for the presented com-
solely certifies the appropriate performance of binational algorithm in this paper. According
deep learning algorithms in the classification to Table 1, this value was acceptable, in com-
of the desired images. The results show that parison with another considered algorithm.
ResNet101 pre-trained neural network has However, what determines the importance of
obtained the highest performance for AUC, this algorithm is the evaluation error index
equal by 0.9998. The proposed algorithm can of 10-fold cross validation, besides accuracy,
achieve to acceptable performance of AUC, sensitivity, and specificity in Figure 6 and Ta-
equal to 0.9999 as well. The worst AUC was ble 1, the resulted value for the proposed algo-
obtained by Inception-ResNetv2, 0.9757. The rithm is less than each of pre-trained networks
resulted AUC for AlexNet and GoogleNet are alone. In this work, we didn’t weight and train
approximately similar which show that, be- a CNN network from scratch, but we used of
cause of the low number of comprised layers pre-trained networks to fine-tune the weights
constructing their networks, these networks of different layers. To clarify the performance
show almost the same outcome, in their AUCs. of the proposed algorithm, Table 2 presents
On the other hand, using of 10-fold cross vali- the outcome of the various evaluation in com-
dation algorithm on the proposed algorithm parison with the proposed algorithm. It shows
shows the lowest error in comparison with us- well the strength of the proposed algorithm in
ing the networks independently. accuracy and sensitivity, although its lack in
Therefore, deep learning algorithms provide the dataset.
appropriate results on evaluation and classifi- Along with all interpretations, it can be sum-
cation approaches for liver hepatocytes with marized that classification by deep learning
fat infiltration level. Although information for algorithms can be considered as an efficient
55 different patients in 10 image sequences method than previous experimental manners.

Table 2: Benchmarking table

Dataset Classifier Accuracy Sensitivity Specificity

Authors Features
size type (%) (%) (%)
Byra [19] 550 Inception-ResNetV2 features SVM 96.3 100 88.2
Stacked sparse Auto-encoder-
Hassan [15] 110 SoftMax 97.2 98 95.7
based features
Kuppili [14] 63 GLCM-based features SVM 86.42 88.20 86.30
Reddy [23] 1000 CNN-based features SVM 93.5 95.3 96.68
VGGNet, ResNet, GoogleNet-
Khan [28] 8000 SoftMax 97.52 97.5 N/A
based features
Birjandi [29] 1600 Clinical factors CT 80 74 83
ResNetV2, GoogleNet,
Proposed
550 AlexNet, and ResNet101- SVM 98.64 97.20 100
Framework
based features
GLCM: Gray-Level Co-Occurrence Matrix, CNN: Convolutional neural network, VGGNet: Visual Geometry Group Network,
SVM: Support vector machine, CT: Classification Tree

J Biomed Phys Eng 2021; 11(1) 81

Zamanian H. et al
Because using this course is followed by sev- Diseases, American College of Gastroenterol-
eral advantages, first of all, there is no need to ogy, and the American Gastroenterological As-
have an interaction with an expert and it can sociation. Hepatology. 2012;55(6):2005-23. doi:
classify the acquired information automati- 10.1002/hep.25762. PubMed PMID: 22488764.
cally. In other words, assigning an ROI region 2. Koncarevic A, Lachey J, Seehra J, Sherman
ML., inventors. Acceleron Pharma Inc, assignee.
is not necessary. Secondly, in this work, we
Methods for treating fatty liver disease. US pat-
employed several images collections for train- ent. 15/148,162. 2016 Nov 3.
ing the network for classification, whilst in the
3. Litjens G, Kooi T, Bejnordi BE, Setio AA, et al.
previous methods, classification was done by A survey on deep learning in medical image
single images and there was no comprehen- analysis. Medical Image Analysis. 2017;42:60-
sive way. 88. doi: 10.1016/j.media.2017.07.005. PubMed
The present paper sustains some limitations PMID: 28778026.
that need to be addressed. First, we utilized 4. Ganapathy N, Swaminathan R, Deserno TM.
a dataset presented in [19] and they couldn’t Deep learning on 1-D biosignals: a taxonomy-
be validated by authors. Second, the limita- based survey. Yearbook of Medical Informatics.
tion of the database place constraints on the 2018;27(1):98. doi: 10.1055/s-0038-1667083.
PubMed PMID: 30157512. PubMed PMCID:
performance of our proposed algorithm, and it PMC6115218.
can be improved by accumulating with other
5. Vanderbeck S, Bockhorst J, Komorowski
patients’ information. Apart from that, we em- R, Kleiner DE, Gawrieh S. Automatic clas-
ployed k-fold cross validation which is suit- sification of white regions in liver biopsies
able for small size datasets. Although, it shows by supervised machine learning. Human Pa-
more bias in prediction error that it is not ap- thology. 2014;45(4):785-92. doi: 10.1016/j.hum-
propriate by increasing data size. path.2013.11.011. PubMed PMID: 24565203.
6. Hashem EM, Mabrouk MS. A Study of Sup-
Conclusion port Vector Machine Algorithm for Liver Dis-
In this paper, a new approach for the imple- ease Diagnosis. American Journal of Intelli-
gent Systems. 2014;4(1):9-14. doi: 10.5923/j.
mentation of pre-trained CNN was presented ajis.20140401.02.
for the classification of the acquired ultrasound
7. Rau HH, Hsu CY, Lin YA, Atique S, et al. Devel-
images of the liver textures. The propounded opment of a web-based liver cancer prediction
approach demonstrates the performance and model for type II diabetes patients by using an
independence of the method from the user or artificial neural network. Computer Methods and
related expert interference. In this approach, Programs in Biomedicine. 2016;125:58-65. doi:
the results of different deep learning algo- 10.1016/j.cmpb.2015.11.009. PubMed PMID:
rithms were presented and compared, based on 26701199.
their performances. The results of this study 8. Hinton G. Deep learning—a technology
show that proposed pre-trained CNN can be with the potential to transform health care.
Jama. 2018;320(11):1101-2. doi: 10.1001/
used with high accuracy to classify ultrasound
jama.2018.11100. PubMed PMID: 30178065.
images of the liver to normal or fatty.
9. Cao GT, Shi PF, Hu B. Liver fibrosis identifica-
tion based on ultrasound images captured un-
Conflict of Interest der varied imaging protocols. J Zhejiang Univ
None Sci B. 2005;6(11):1107. doi: 10.1631/jzus.2005.
B1107. PubMed PMID: 16252346. PubMed PM-
References CID: PMC1390659.
1. Chalasani N, Younossi Z, Lavine JE, Diehl AM, 10. Acharya UR, Faust O, Sree SV, Molinari F, Suri
et al. The diagnosis and management of non-al- JS. ThyroScreen system: high resolution ultra-
coholic fatty liver disease: Practice Guideline by sound thyroid image characterization into benign
the American Association for the Study of Liver and malignant classes using novel combination

82 J Biomed Phys Eng 2021; 11(1)

 Combinational Deep Learning Algorithm for NAFLD Classification
of texture and discrete wavelet transform. Com- 19. Byra M, Styczynski G, Szmigielski C, Kalinowski
put Methods Progr Biomed. 2012;107(2):233- P, et al. Transfer learning with deep convolu-
41. doi: 10.1016/j.cmpb.2011.10.001. PubMed tional neural network for liver steatosis as-
PMID: 22054816. sessment in ultrasound images. Int J Comput
11. Andrade A, Silva JS, Santos J, Belo-Soares P. Ass Rad. 2018;13(12):1895-903. doi: 10.1007/
Classifier approaches for liver steatosis using s11548-018-1843-2. PubMed PMID: 30094778.
ultrasound images. Proc Tech. 2012;5:763-70. PubMed PMCID: PMC6223753.
doi: 10.1016/j.protcy.2012.09.084. 20. Marshall RH, Eissa M, Bluth EI, Gulotta PM,
12. Gao S, Peng Y, Guo H, Liu W, Gao T, Xu Y, Tang Davis NK. Hepatorenal index as an accurate,
X. Texture analysis and classification of ultra- simple, and effective tool in screening for ste-
sound liver images. Bio-Medical Materials and atosis. Am J Roentgenol. 2012;199:997-1002.
Engineering. 2014;24(1):1209-16. doi: 10.3233/ doi: 10.2214/AJR.11.6677. PubMed PMID:
BME-130922. PubMed PMID: 24212015. 23096171.
13. Acharya UR, Raghavendra U, Fujita H, Hagi- 21. Wu CC, Yeh WU, Hsu WD, Islam MM, et al.
wara Y, et al. Automated characterization of Prediction of fatty liver disease using ma-
fatty liver disease and cirrhosis using curvelet chine learning algorithms. Comput Methods
transform and entropy features extracted from Progr Biomed. 2019;170:23-9. doi: 10.1016/j.
ultrasound images. Computers in Biology and cmpb.2018.12.032. PubMed PMID: 30712601.
Medicine. 2016;79:250-8. doi: 10.1016/j.comp- 22. Mohaimenul Islam M, Wu CC, Poly TN, Yang
biomed.2016.10.022. HC, Li YC. Applications of machine learning in
14. Kuppili V, Biswas M, Sreekumar A, Suri HS, fatty live disease prediction. Stud Health Technol
Saba L, et al. Extreme learning machine frame- Inform. 2018;247:166-70. doi: 10.3233/978-1-
work for risk stratification of fatty liver disease 61499-852-5-166. PubMed PMID: 29677944.
using ultrasound tissue characterization. Jour- 23. Reddy DS, Bharath R, Rajalakshmi P. A novel
nal of Medical Systems. 2017;41(10):152. doi: computer-aided diagnosis framework using deep
10.1007/s10916-017-0797-1. PubMed PMID: learning for classification of fatty liver disease in
28836045. ultrasound imaging. International Conference on
15. Hassan TM, Elmogy M, Sallam ES. Diagnosis of e-Health Networking, Applications and Services
focal liver diseases based on deep learning tech- (Healthcom); Ostrava, Czech Republic: IEEE;
nique for ultrasound images. Arabian Journal for 2018. doi: 10.1109/HealthCom.2018.8531118.
Science and Engineering. 2017;42(8):3127-40. 24. Frid-Adar M, Klang E, Amitai M, Goldberger J,
doi: 10.1007/s13369-016-2387-9. Greenspan H. Synthetic data augmentation us-
16. Liu X, Song J, Wang S, Zhao J, Chen Y. Learn- ing GAN for improved liver lesion classifica-
ing to diagnose cirrhosis with liver capsule tion. international symposium on biomedical
guided ultrasound image classification. Sen- imaging (ISBI). 2018;289-93. doi: 10.1109/
sors. 2017;17(1):149. doi: 10.3390/s17010149. ISBI.2018.8363576.
PubMed PMID: 28098774. PubMed PMCID: 25. Kleiner DE, Brunt EM, Van Natta M, Behling C, et
PMC5298722. al. Design and validation of a histological scor-
17. Bharath R, Rajalakshmi P. Deep scattering con- ing system for nonalcoholic fatty liver disease.
volution network-based features for ultrasonic Hepatology. 2005;41(6):1313-21. doi: 10.4196/
fatty liver tissue characterization. Annual Inter- kjpp.2019.23.6.459. PubMed PMID: 31680767.
national Conference of the IEEE Engineering in PubMed PMCID: PMC6819900.
Medicine and Biology Society (EMBC); Seog- 26. Perez L, Wang J. The effectiveness of data aug-
wipo, South Korea: IEEE; 2017. p. 1982-5. doi: mentation in image classification using deep
10.1109/EMBC.2017.8037239. learning. arXiv preprint arXiv:1712.04621. 2017.
18. Biswas M, Kuppili V, Edla DR, Suri HS, Saba L, 27. Kinnunen T, Li H. An overview of text-independent
et al. Symtosis: A liver ultrasound tissue char- speaker recognition: From features to supervec-
acterization and risk stratification in optimized tors. Speech Communication. 2010;52(1):12-
deep learning paradigm. Comput Methods Progr 40. doi: 10.1016/j.specom.2009.08.009.
Biomed. 2018;155:165-77. doi: 10.1016/j. 28. Khan S, Islam N, Jan Z, Din IU, Rodrigues JJ.
cmpb.2017.12.016. PubMed PMID: 29512496. A novel deep learning-based framework for

J Biomed Phys Eng 2021; 11(1) 83

Zamanian H. et al
the detection and classification of breast can- holic fatty liver disease (NAFLD) and identification
cer using transfer learning. Pattern Recogni- of its associated factors using the classification
tion Letters. 2019;125:1-6. doi: 10.1016/j.pa- tree method. Iranian Red Crescent Medical
trec.2019.03.022. Journal. 2016;18(11):e32858. doi: 10.5812/ir-
29. Birjandi M, Ayatollahi SM, Pourahmad S, Safar- cmj.32858. PubMed PMID: 28191344. PubMed
pour AR. Prediction and diagnosis of non-alco- PMCID: PMC5292777.

84 J Biomed Phys Eng 2021; 11(1)