Automatic Classification of Diffuse Liver Diseases:

Cirrhosis & Hepatosteatosis

using Ultrasound Images
Ahmed Ossama, Omar Ahmed, Marwan Hashem,
2022 2nd International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC) | 978-1-6654-6677-6/22/$31.00 ©2022 IEEE | DOI: 10.1109/MIUCC55081.2022.9781734

Mostafa Ahmed, Verina Saber, Alaa Hamdy

Faculty of Computer Science
Misr International University
Cairo, Egypt
ahmed1803760, omar1803098, marwan1805486,
mostafa1806150,verina.saber, alaa.hamdy {@miuegypt.edu.eg}

Abstract The alternative non-invasive techniques that have been

This research aims to develop an intelligent system employed are medical imaging such as Ultrasound,
that can detect and classify various liver diseases using Computerized Tomography scan (CT-scan) and Magnetic
machine learning. This study will employ digital Ultra- Resonance Imaging (MRI), which help medical professionals
sound images (US) collected from various hospitals and identify these diseases without the need for invasive
clinics in Egypt, as it is identified as one of the most
procedures. The MRI continues to be a powerful imaging tool
common and affordable methods of medical imaging
in the country [1], to identify liver diseases in a non- that can be used to accurately diagnose diffuse liver diseases,
invasive manner. We will be extracting multiple Gray- but the accuracy and quality of the imaging results in a very
Level Correlation Matrix (GLCM), Gray Level Run expensive process. The rising prices of MRI Scans lead to
Length Matrix (GLRLM) and First Order Statistical many people in rural areas avoiding processes such as MRI
(FOS) features from each US. Multiple feature selec-
scans as they can rarely afford such a luxury.
tion methods will be used to select the most important
features for each of the 3 binary classifiers. The trained
classifiers achieved an accuracy of 89.7%, 84.2% and When it comes to affordable imaging techniques that
95.5% for the Normal/Steatosis, Normal/Cirrhosis and still produce quality results, US imaging reigns supreme
Steatosis/Cirrhosis classifiers respectively. A majority across the board. Through US imaging, the proposed system
function was used to classify the US based on the result
will be able to classify US images into three distinct classes;
of the 3 binary classifiers, the total accuracy of the
model was 85.3%. Healthy Liver, Steatotic Liver, and Cirrhotic Liver. However,
US imaging itself poses a problem of producing low-quality
Index Terms images, which makes the tasks of feature selection and
Liver Diseases, Machine Learning, Ultrasound, extraction more difficult.
Region of Interest, Feature Extraction

I. I NTRODUCTION
Cirrhosis has been common in Egypt for decades, making it
one of the world’s largest countries with the highest cirrhosis
rates. [2]. Health-care resource constraints exacerbate the
problem by enabling disease-specific consequences to emerge
in the absence of effective therapies in the past, leading Fig. 1. The pathogenesis of liver disease inside the human body
to the consequences of an untreated diseased liver, as Fig.
1 illustrates. The need to address these important health
II. R ELATED W ORK / L ITERATURE S URVEY
issues in Egypt prompted a great deal of hepatology research.
Needle biopsies have been historically, and to this day remain, A. Machine Learning
the go-to method of accurately identifying liver disease [3], One study [4] describes a novel approach for detecting
but they are still an invasive and dangerous procedure. liver abnormalities using ultrasound images automatically.
A machine learning model was developed that can not only
generate labels (normal and abnormal) for an ultrasound
image, but also recognise when the prediction is likely to be
inaccurate. The suggested model does not create the label
978-1-6654-6677-6/22/$31.00 ©2022 IEEE of a test example if it is not confident in its prediction. On

246
Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on February 06,2025 at 08:34:29 UTC from IEEE Xplore. Restrictions apply.
 a dataset of 99 US images, the AUC-ROC of the K-Nearest B. Deep Learning
Neighbour (KNN) classifier was 78%, while the AUC-ROC
of the conventional Support Vector Machine (SVM) was 87%. An alternative study [7] uses a feature-based approach
instead of higher order local auto-correlation (HLAC), which
Another approach [5] intends to improve the effectiveness has previously shown unsatisfactory results, to identify the
of computer-assisted liver disease diagnosis by improving US condition faster using computer aided diagnosis (CAD).
images before classifying them. his study describes a unique Convolutional neural networks (CNNs) are frequently
approach for autonomously extracting the liver capsule, as employed in medical image diagnosis, however, they have the
well as an algorithm that uses a pre-trained SVM classifier to drawback of requiring a large number of training samples, as
extract features of the parenchyma to determine the degree of well as over-training due to the limited amount of available
pathological alterations. On a dataset of 48 cirrhotic liver & examples. On a dataset of 500 US images (200 normal liver,
20 normal liver US images, the suggested technique attained 300 cirrhotic liver), the perspective augmentation achieved
an accuracy of 92.08%, 80.23%, 75.12%, and 93.58% for the greatest results with an error rate of 31.2% compared to
cases with normal, mild, moderate, and severe cirrhosis, an error rate of 37.7% without augmentation.
respectively, using a k-fold cross-validation strategy with K
set to 5. An additional approach [8] offers an end-to-end deep
learning strategy for automatic cirrhosis classification
using ultrasound images. To support region of interest
Furthermore, another study [6] aims to improve
(ROI) extraction from various ultrasound images, automatic
classification of normal and cirrhotic liver using various
ROI detection module was included. Simultaneously, the
texture features and weighted fisher discriminant classifier.
classification module employs ROI areas and the transfer
The Fisher Discriminant Ratio (FDR) is utilised to reduce the
learning network to obtain cirrhosis diagnosis results. On a
number of features in this investigation. The features with a
dataset of 681 samples, LiverTL had an AUC of over 0.90.
higher FDR value have greater discriminative strength and
are thus chosen for classification using the established FDR
method. On a dataset of 60 cirrhotic liver & 72 normal liver Some further studies [11] used the singular value
US images, the model resulted in a high accuracy of 96.38% decomposition (SVD) of the GLCM matrix, to classify
and a sensitivity of 98.33%, this algorithm can differentiate between normal and cirrhotic liver ultrasounds. They reached
cirrhotic liver from normal liver. an accuracy of 95.04% in both cases, as they combined the
SVD with a Neural Network (NN) on 82 normal ROIs and
39 cirrhotic ROIs from 12 patients.
Moreover, some other approach [9] proposed a system
to distinguish between normal, cirrhotic, and steatotic liver
disorders. This method combines two feature selection Finally an alternative approach [12] proposed a collection
methods, avoiding their drawbacks while maximising their of classifiers (SVM, Naive Bayes, Parzen, KNN) to get the
benefits, and bases categorization on the decisions of three best result possible, and the best result was achieved using
sub-classifiers. On a dataset of 279 US images (95 normal, KNN where K = 1 and the results were 97% and 99% for the
105 steatotic, and 79 cirrhotic), the recognition accuracy of normal and steatotic liver respectively. The study was done
the normal/steatosis classifier was 95%, the normal/cirrhosis on 75 ultrasounds of 75 patients with 40 of them having a
classifier was 95.74%, and the steatosis/cirrhosis classifier normal liver and 35 having steatoic liver.
was 94.23%. To arrive at a final conclusion of 95% accuracy,
the majority function is applied.
III. M ATERIALS AND M ETHODS
An additional study [10] proposed the AdaBoost approach
to combine weak classifiers and the use of an artificial dataset A. Image Acquisition
to overcome the challenge of overfitting with the relatively
small dataset of 20 ultrasound images, which bettered the A dataset was constructed from US images obtained from
original results from 56% to 80%, using cross validation. multiple sources and local clinics, thus there isn’t a specific
US machine being considered. A total of 61 images were
An alternative study [13] approached the problem made available by Ahmed Gaber Mabrouk [9], 55 images
of classifying healthy and cirrhotic liver, where they were obtained from a public dataset [14] an image from each
experimented using SVM classifier with 3 different feature patient in the study was used, and 73 images from various local
selection methods(PCA - u-LBP - GLCM) with the best clinics. A total of 189 US images were taken from different
accuracy to the u-LBP-SVM which was 87%. The dataset patients, and were manually classified as normal, steatotic, or
used for this research consisted of 20 normal images and 40 cirrhotic by medical professionals, 70 images were classified
cirrhotic ones. as normal, 94 images were classified as steatotic and 25 were
classified as cirrhotic as shown in Fig. 2.

247
Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on February 06,2025 at 08:34:29 UTC from IEEE Xplore. Restrictions apply.
 of the ROIs cannot be fixed,and all the ROIs selected were
manually marked by medical professionals.

Furthermore, a total of 108 GLRLM, GLCM, and FOS

features [15] were extracted from each ROI as well as the
diagonal length of the liver. GLCM features were calculated
with distances 1, 2 and 3 pixels, as well as angles 0, 45,
Fig. 2. Different States of the Liver using Ultrasound Imaging 90, and 135, Each feature was calculated 12 times with each
distance and angle.
B. Methodology
To improve the performance of the trained classifiers Finally, feature selection was conducted using each of
multiple ROIs are identified inside the liver tissue rather than Fisher’s score (F-ratio), the Random Forest Classifier’s feature
a single ROI, as shown in Fig. 3. A size of 32x32 pixels was importance, and manual feature selection. Manual feature
chosen for all ROIs [6]. Each image contains a minimum of selection was achieved by training a model using a single
6 ROIs and a maximum of 20 depending on the image. A feature and the model’s accuracy was used as the feature’s
total of 2257 ROIs were extracted from all US images in the score. Each feature selection method was used on each binary
dataset. Since the position and size of the liver as well as the classifier. The feature selection task was carried out to select
angle at which the image was captured and the size of the the most prominent features, in order to increase the model’s
image are not constant throughout the dataset, the position accuracy when classifying the US images.

Fig. 3. System Overview illustrating the flow of the proposed system

IV. R ESULTS AND D ISCUSSION (ANN) classifier. The total accuracy of the model was 75.0%
with no abstentions.
Firstly, the study attained slightly biased results, owing to
the skew in sample size between normal, cirrhotic and steatotic
images, the skew being a substantial increase in steatotic
images compared to normal and cirrhotic images, as well as a
substantial decrease in cirrhotic images compared to steatotic
images.
Secondly, several experiments were performed throughout
each development phase. First, without performing any feature Third, after utilising different feature selection methods such
selection methods, as shown in Fig. 4, the model produced as Random Forest, Fisher’s score and manual feature selection
an accuracy of 81.3% for the Normal/Steatosis classifier, to compare their results, Manual feature selection produced
70.0% for the Normal/Cirrhosis classifier, and 90.9% for the the best results using the ANN classifier. The most important
Steatosis/Cirrhosis classifier using Artificial Neural Network features for each classifier are as shown in Fig. 5.

248
Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on February 06,2025 at 08:34:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 6. Best number of features for the 3 classifiers using manual feature
Fig. 4. The improvement of the model’s accuracy after feature selection and selection
adding the length feature
As shown in Fig. 7, for fisher score feature selection,
the number of features to output the best results were 12
features for the Normal/Steatosis clasifier, 94 features for the
Normal/Cirrhosis classifier and 62 features for the Steato-
sis/Cirrhosis classifier.

Fig. 5. Top features using manual feature selection

As shown in Fig. 4, by using ANN classifier,

manual feature selection produced 82.8% accuracy for
Normal/Steatosis classifier using 21 features, 78.9% accuracy
for Normal/Cirrhosis classifier using 82 features, and 95.5%
accuracy for Steatosis/Cirrhosis classifier using 6 features. Fig. 7. Best number of features for the 3 classifiers using Fisher’s Score
feature selection
The total accuracy of the model was 76.4% with an abstention
rate of 2.8%.
As shown in Fig. 8, for random forest feature selection,
the number of features to output the best results were 12
Fourth, we were able to add a new feature for each features for the Normal/Steatosis clasifier, 7 features for the
US image in the dataset, which is the liver’s length by Normal/Cirrhosis classifier and 76 features for the Steato-
measuring the midclavicular line. The length was calculated sis/Cirrhosis classifier.
by measuring the maximum Euclidean distance between the
ROIs located on either diagonal of the liver in the US image.
The most prominent features extracted are shown in Fig. 5.

After extracting all of the available features from the US

images and discovering the most prominent features, it was
important to determine the exact number of important features
that output the best accuracy. Therefore, an algorithm was de-
veloped to identify the exact number of features each classifier
requires to return the greatest results. As shown in Fig. 6, for
manual feature selection; the number of features to output the
best results were 20 features for the Normal/Steatosis clasifier,
7 features for the Normal/Cirrhosis classifier and 25 features Fig. 8. Best number of features for the 3 classifiers using Random Forest
for the Steatosis/Cirrhosis classifier. feature selection

249
Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on February 06,2025 at 08:34:29 UTC from IEEE Xplore. Restrictions apply.
 As shown in Fig. 4, after adding the liver length feature, [9] A. G. Mabrouk, A. Hamdy, H. M. Abdelaal, A. G. Elkattan, M. M.
the model, using the ANN classifier, was able to produce an Elshourbagy, and H. A. Alansary, “Automatic classification algorithm
for diffused liver diseases based on ultrasound images,” IEEE Access,
accuracy of 89.7% for the Normal/Steatosis classifier using vol. 9, pp. 5760–5768, 2021.
20 features, 84.2% for the Normal/Cirrhosis classifier using [10] Y. Fujita, Y. Hamamoto, M. Segawa, S. Terai, and I. Sakaida, “An
7 features and 95.5% for the Steatosis/Cirrhosis classifier improved method for cirrhosis detection using liver’s ultrasound im-
ages,” 2010 20th International Conference on Pattern Recognition, pp.
using 25 features. The total accuracy of the model was 85.3% 2294–2297, 2010.
with an abstention rate of 2.8%. The ANN contained two [11] J. Virmani, V. Kumar, N. Kalra, and N. Khandelwal, “Prediction of
hidden layers and used the ADAM solver [16] and the RELU cirrhosis based on singular value decomposition of gray level co-
occurence Marix and Anneural network classifier,” 2011 Developments
activation function. in E-systems Engineering, pp. 146–151, 2011.
[12] R. T. Ribeiro, R. T. Marinho, and J. M. Sanches, “Classification and
staging of chronic liver disease from multimodal data,” IEEE Transac-
tions on Biomedical Engineering, vol. 60, no. 5, pp. 1336–1344, 2013.
[13] Y.-ming Lei, X.-mei Zhao, and W.-dong Guo, “Cirrhosis recognition of
V. C ONCLUSION liver ultrasound images based on SVM and uniform LBP feature,” 2015
IEEE Advanced Information Technology, Electronic and Automation
Individual distinctions are made between normal/steatosis, Control Conference (IAEAC), pp. 382–387, 2015.
normal/cirrhosis, and steatosis/cirrhosis. Normal/steatosis clas- [14] M. Byra, G. Styczynski, C. Szmigielski, P. Kalinowski, Ł. Michałowski,
R. Paluszkiewicz, B. Ziarkiewicz-Wróblewska, K. Zieniewicz, P. So-
sifier had a recognition accuracy of 89.7%, normal/cirrhosis bieraj, and A. Nowicki, “Transfer learning with deep convolutional
classifier had a recognition accuracy of 84.2%, and steato- neural network for liver steatosis assessment in ultrasound images,”
sis/cirrhosis classifier had a recognition accuracy of 95.5%. International Journal of Computer Assisted Radiology and Surgery, vol.
13, no. 12, pp. 1895–1903, 2018.
The majority function is used to make the final decision. The [15] J. J. M. van Griethuysen, A. Fedorov, C. Parmar, A. Hosny, N. Aucoin, V.
attained accuracy of recognition is 85.3% with an abstention Narayan, R. G. H. Beets-Tan, J.-C. Fillion-Robin, S. Pieper, and H. J. W.
rate of 2.8%. L. Aerts, “Computational radiomics system to decode the radiographic
phenotype,” Cancer Research, vol. 77, no. 21, pp. 104–107, Nov. 2017.
The proposed approach can be utilised to distinguish between [16] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization.”
the different stages of diffused liver fibrosis, such as classify- arXiv, 2014.
ing US image as either a normal, steatotic or cirrhotic liver.

ACKNOWLEDGMENT
We would like to thank Christine Raouf Zaki, radio diag-
nosis consultant, Ahmed Gaber Mabrouk [9] for their contri-
bution to our dataset.

R EFERENCES
[1] T. Mansour, M. Tawfik, M. El-Barody, S. Sileem, and A. Okasha,
“Correlation between ultrasound and magnetic resonance imaging in
diagnosis of ovarian tumors,” Al-Azhar International Medical Journal,
vol. 1, no. 12, pp. 214–223, Dec. 2020.
[2] A. Gomaa, N. Allam, A. Elsharkawy, M. El Kassas, and I. Waked,
“Hepatitis C infection in Egypt: Prevalence, impact and management
strategies,” Hepatic Medicine: Evidence and Research, vol. Volume 9,
pp. 17–25, May 2017.
[3] F. M. Sanai and E. B. Keeffe, “Liver biopsy for histological assessment
- the case against,” Saudi Journal of Gastroenterology, vol. 16, no. 2,
pp. 124–132, Mar. 2010.
[4] K. Hamid, A. Asif, W. Abbasi, D. Sabih, and F.-ul-A. A. Minhas, “Ma-
chine learning with abstention for automated liver disease diagnosis,”
2017 International Conference on Frontiers of Information Technology
(FIT), pp. 356–361, 2017.
[5] X. Liu, Z. Zhan, M. Yan, J. Zhao, J. Song, and Y. Q. Chen, “Computer-
aided cirrhosis diagnosis via automatic liver capsule extraction and com-
bined geometry-texture features,” 2017 IEEE International Conference
on Multimedia and Expo (ICME), pp. 865–870, 2017.
[6] S. K. Randhawa, R. K. Sunkaria, and A. K. Bedi, “Prediction of liver
cirrhosis using weighted fisher discriminant ratio algorithm,” 2018 First
International Conference on Secure Cyber Computing and Communica-
tion (ICSCCC), pp. 184–187, 2018.
[7] Y. Mitani, R. B. Fisher, Y. Fujita, Y. Hamamoto, and I. Sakaida, “Effect
of an augmentation on cnns in classifying a cirrhosis liver on B-mode
ultrasound images,” 2020 IEEE 2nd Global Conference on Life Sciences
and Technologies (LifeTech), pp. 253–254, 2020.
[8] H. Yang, X. Sun, Y. Sun, L. Cui, and B. Li, “Ultrasound image-
based diagnosis of cirrhosis with an end-to-end deep learning model,”
2020 IEEE International Conference on Bioinformatics and Biomedicine
(BIBM), pp. 1193–1196, 2020.

250
Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on February 06,2025 at 08:34:29 UTC from IEEE Xplore. Restrictions apply.