Page 1353
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
A Review of Artificial Intelligence-Based Approaches for Non-
Invasive Liver Disease Diagnosis
Srishti Bhardwaj
1
, Sonam Aggarwal
2
, Abhilasha Sood
1
, Abhishek Prasad
3
1
Chitkara School of Health Sciences, Chitkara University, Punjab, India.
2
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India.
3
Department of Radiology, Fortis Hospital, Mohali
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.15020000120
Received: 28 February 2026; Accepted: 05 March 2026; Published: 23 March 2026
ABSTRACT
Liver disease is a significant clinical problem on a global scale and a leading cause of morbidity and mortality.
Early and prompt diagnosis is essential to successful management, but the traditional diagnosis methods tend to
require invasive procedures, or they are restricted by the limitations of varying observers. In recent years, the
use of artificial intelligence (AI) and computer-aided diagnosis (CAD) systems has acquired significant
popularity in the field of hepatology. Machine learning (ML) and deep learning (DL) technologies are actively
used in medical imaging data, including ultrasound, CT, and MRI, to assist clinicians in identifying and
classifying liver pathologies, i.e., fatty liver disease, cirrhosis, and hepatocellular carcinoma. The review is a
synthesis of the recent developments in image processing, feature detection, and classification algorithms of
liver disease diagnosis by AI. The most common ML algorithms include Support Vector Machines, random
forests, decision trees, naive bayes, and K-nearest neighbors, in many cases using radiomic features derived
using imaging data. Although deep learning models, especially convolutional neural networks and transfer
learning implementations, are highly sensitive and highly perform in segmentation and classification tasks,
traditional ML systems with radiomic features are frequently able to offer robust and efficient solutions to
resource-bound environments. Even though these results are promising, there are still a number of challenges
such as data heterogeneity, insufficient multi-center validation, and model interpretability. To reliably translate
clinical findings into clinical practice and enhance patient outcomes, future research should focus on large-scale
validation studies, multimodal data integration, and explainable AI frameworks.
Keywords: Artificial intelligence, Liver Disease Diagnosis, Machine Learning, Deep Learning, Medical
Imaging, Computer-Aided Diagnosis, Radiomics, Image Processing
INTRODUCTION
Liver is an important organ and it plays a characteristic role in many physiological functions including
metabolism, detoxification, immune regulation, energy storage, blood coagulation etc [1]. In spite of the
multifunctionality, liver disease has emerged as one of the most significant global health issues considerably
increasing the morbidity and mortality rates. About two million deaths are attributed to liver related disorders
like cirrhosis, viral hepatitis and liver cancer annually. Liver disease is estimated to cause 4 per cent of the total
deaths all over the world and one out of every twenty five deaths is associated with hepatic dysfunction
[2].Although liver disease is one of the major causes of mortality, its actual burden is undervalued. In terms of
population, India has a high percentage of the disease burden among the affected populations, with 18.3% of all
deaths due to cirrhosis-related illnesses, and China leading with 11% [3].In Europe and America, where the
disease is the primary risk, the use of alcohol is the leading factor of liver disease [4]. The total number of people
around the globe living with chronic hepatitis B (HBV) is 400 million, and over 170 million people continue to
live with persistent hepatitis C (HCV) [2]. An increase in incidence of liver disease with the rise of the Metabolic
Dysfunction Associated Steatotic Liver Disease (MASLD) that was previously referred to as Non Alcoholic
Fatty Liver Disease (NAFLD) has also been observed. The prevalence of MASLD is a stimulus of liver cirrhosis
Page 1354
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
and hepatocellular carcinoma (HCC) development with a high incidence in the regions with high obesity rates
such as India. Estimates of the number of overweight and obese adults are approximately 2 billion and 400
million adults have diabetes, both significant risk factors of MASLD and HCC. Also, in spite of high global
burden of hepatitis caused by viruses, drug induced liver injury is now identified as a key cause of acute hepatitis.
[3]. Not only is chronic liver disease a clinical issue, but also an economic burden and quality of life determinant.
In terms of global and regional DALYs and years of life lost, cirrhosis is always in the top 20 causes of health
related disabilities. The common and intricate liver diseases require correct and readily accessible diagnostic
procedures to enhance the patient outcome and lower healthcare expenses.
The modern histopathological and imaging systems have transformed the diagnosis of liver diseases and their
integration has helped in the improvement of patient care, characterisation of the disease and proper diagnosis.
Liver biopsy, which has continued to be an anchor to the diagnosis, is ordered especially in cases where there is
a failure in imaging to provide a clarifying outcome. Diagnostic accuracy has also been provided by
immunohistochemistry of particular marker associated with liver tumors. However, because of the invasiveness
of the biopsy, there has been a massive interest in other less invasive and more scalable methodologies [5].The
need to develop less invasive and more scalable techniques has led to advances in imaging and computer
analysis. Ultrasound (US) has always served as the front-line imaging technique because it is easy to access,
affordable and noninvasive [6]. It remains an important method of diagnosing conditions, such as cirrhosis, fatty
liver disease and hepatomegaly. Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) became
well-established as a more sensitive technique due to structural analysis as US has low-resolution of smaller
abnormalities [7]. The CT images are required in medical personnel when assessing focal liver lesions, staging
and planning surgery of the HCC due to its high spatial resolution and ability to produce a detailed cross-sectional
image [8]. The
use of contrast on CT imaging of growths allows medical personnel to differentiate between
benign and malignant growths that enhances diagnostic decisions. The MRIs offer superior soft-tissue contrast
that helps the professionals detect different masses in the liver with high efficiency [9]. The level of expertise
attained by the observer is very important in interpreting radiological images and hence results in variability of
interpretations. Liver imaging has been rapidly evolving due to the ability of radiomics technology to extract
quantitative characteristics of any medical image and provide accurate diagnosis of liver diseases such as fibrosis
and steatosis. Upon analysis by ML algorithms these extracted features allow imaging to predict more efficiently,
offering clinicians useful information about disease progression and responses to treatment. Studies indicate that
the radiomic analysis has an accuracy of 95.98% in detecting the following elements found in the liver as far as
classification of liver diseases is concerned; ballooning is essential in liver disease diagnosis [10].
The introduction of artificial intelligence (AI) into liver imaging is used to decrease the variability of the
observers and increase diagnostic accuracy. The radiomics-extracted imaging characteristics allow AI-based
Computer Aided Diagnosis to obtain superior liver pathology classification. ML and DL technologies used in
CAD allow making the diagnosis of liver diseases more accurate by providing better workflow management and
minimizing human error. Convolutional neural networks (CNNs) were used and have led to tremendous
advancements in the segmentation of liver images in both MRI and CT imaging to identify fine abnormalities.
Recent research using more intensive convolutional models has built upon these improvements, with the
architecture, e.g. DenseNet 121 and ResNet 50, pushing the accuracy of classification even further higher, with
DenseNet 121 being particularly impressive with its capability to propagate features and gradients across many
layers efficiently [11]. Predictive DL models prove an effective HCC prognosis in patients with steatotic liver
disease and provide precise outcomes with a high accuracy of over 81 percent [12]. Conventional ML methods
can help these new advances by assessing liver functioning test data and picture pattern identification to predict
hepatic issues in order to preventive health early and avoid future health issues. On its part, digital pathology has
demonstrated greater diagnostic accuracy after being integrated with ML since the latter-ledclassifiers have
comparable evaluation capabilities with pathologists when analysing histopathological slides [13]. The growing
popularity of AI-based diagnostic tools suggests that radiomics with CAD technology will gain greater
significance in the diagnosis and treatment of liver disorders. The analysis explores the usefulness of ML and
DL models in the context of CAD tools in the diagnosis of liver disease by image processing methods. The article
unveils the advantages of CAD applications in clinical practice that depend on the possibility to reduce variations
in observers and enhance the accuracy of the diagnosis. Relevant literature studies were based on the IEEE
Digital Library and PubMed as well as ScienceDirect based on the key words: liver diseases, medical imaging,
Page 1355
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
and computer-aided diagnosis. Figure 1 shows the process of having selected the papers included in this review.
Scopus, Google Scholar, and PubMed were used to gather 240 articles by selecting the keywords based on
computer-aided diagnosis, machine learning, D, and liver disease. Articles on the use of CAD systems and ML
and DL models in diagnosis and prediction of liver disease were evaluated. A total of 97 articles are incorporated
in this research after passing the inclusion and exclusion criteria.
The perspectives that this review is expected to address are the following:
1. Detailed study of different image processing methods and features analysis used in the diagnosis of different
liver diseases.
2. The state of art liver CAD systems developed on various radiological modalities such as CT, MRI,
Ultrasound are described comparatively.
3. The use of CAD in diagnosis and detection of different liver diseases.
4. The thorough comparison of the various ML and DL models in the classification of multiple class liver
diseases.
The paper makes novel contributions to the CAD of liver disease. It reviews the current research on the data
diversity in CAD models, locating the gaps and ways to enhance the data reliability to increase the performance
of diagnostic measures. It further discusses transparency of CAD systems in medical imaging, discussing the
issues in clarifying model decisions and its consequences on clinical trust and acceptance.
Figure 1. PRISMA flow diagram illustrating the study identification, screening, eligibility, and inclusion process
for the review.
Fundamentals of Computer Aided Diagnosis
CADs have become essential areas of research within the past twenty years with a focus on improving the quality
of diagnosis and facilitating the treatment process of radiologists and clinicians who work on the medical image
analysis. The role of big data and AI, and in particular ML and DL algorithms in the field of healthcare is
groundbreaking as it has made the field of healthcare more efficient, proactive, and personalized. CAD has
become a fundamental discipline of medical imaging and diagnostic radiology. Although the first computerized
Page 1356
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
medical image analysis had been started in the 1960s, the systematic research activities were the ones that gained
momentum in the 1980s. This transition carried a shift in the CAD as a completely automated system towards a
supportive tool that will increase the accuracy of the diagnosis made by the radiologist by giving unbiased data-
driven information [14].At the end of the 1980s and at the beginning of the 1990s, AI-driven approaches,
especially ML and DL methods, started to change the nature of CAD systems, turning them into a second opinion
that will aid in the precision of the diagnostic decision made by the radiologist. These systems are usually
implemented as a sequence of steps that include image processing, feature extraction and classification steps and
tend to rely on artificial neural networks (ANN) to maximize detection and
features [15]. The development of
DL has continued to speed up in CAD and this has seen a big difference in the accuracy and efficiency of
detection. As technologies keep progressing, AI CAD models are transforming radiological imaging, making
way to more efficient and accurate diagnostic models. Figure 2 gives a visual overview of the pipeline used in
various CAD programs and the standard procedures applied.
Figure 2: Pipeline architecture of a CAD system, detailing key stages from image acquisition to classification.
Image Acquisition
Medical imaging is a vital instrument in the early diagnosis of liver diseases and how to define their nature and
track them to enable planning of their treatment
. Both of the diagnostic methods have certain merits and the
demerits that determine their reasonable use in the medical practice. The medical imaging techniques such as
US, CT and MRI are the necessary ones in the diagnosis of liver diseases due to the varied benefits of sensitivity
and specificity and clinical use. This segment explores various imaging procedures applied in the diagnosis of
liver pathology and explains their advantages and limitations to the medical practice.
Ultrasound
The US examination of the liver is a critical part of disease diagnosis as it provides non-invasive and effective
early detection facets as well as monitoring qualities. The recent advancements in imaging technology have
enhanced the diagnosis of liver conditions by multiparametric ultrasound and elastography techniques that have
enhanced the detection of simple steatosis and complex vascular diseases. The advancement of the new
technology enables the doctor to thoroughly examine the liver anatomy and functions that result in improved
clinical judgment. B-mode ultrasound preserves its original status of measuring liver size and measuring
echogenicity whereas Doppler ultrasound identifies liver diseases by measuring blood flow [16]. Elastography
not only allows the measurement of liver stiffness to quantify it, but also serves as an alternative measure to
fibrosis in the monitoring of the progression of chronic liver disease [17,18]. Findings in research indicate that
ultrasound is very sensitive in detecting the presence of abnormalities in the liver with particular attention to
MASH [16,17]. Ultrasound has been more successful in diagnosis of diffuse liver parenchymal diseases when
used with other imaging techniques such as FibroScan [18].
Page 1357
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Nevertheless, there are a number of intrinsic ultrasound drawbacks that need to be taken into consideration. Bone
and air-filled structures to a great extent block the passage of ultrasound waves, so that it is not possible to view
areas that are covered by the ribs, bowel gas, or lungs. Furthermore, the differences in tissue composition within
the abdomen cause inhomogeneity in the beam which scatters and bends the beam and consequently ruins the
sharpness of the image in practice and diffraction-limited resolution is impractically achieved. Ultrasound also
has poor sensitivity in mild steatosis instances and therefore it may require supplementary diagnostic studies to
be conducted as a whole [19]. Moreover, its diagnostic reliability is highly subjective, and thus there is
inconsistency in the outcomes unlike liver biopsy where the histological confirmation is essential and therefore
the use of other diagnostic methods will be essential. The future of ultrasound in assessing liver disease is in the
optimization of the ultrasound imaging modalities, the addition of artificial intelligence, and the strategic
integration of ultrasound with other modalities to create a more precise and global diagnostic model.
Computed Tomography
CT imaging is a reliable method of detecting and assessing liver pathologies due to the fact that it provides key
anatomical and functional data required in clinical decisions. These high-resolution imaging characteristics assist
physicians with determining liver tumors at an exact localization and also characterize the tumor beside crucial
volumetric measurements to assist in preoperative planning. The CT elucidates the feasibility of the surgery by
measuring the total liver volume and tumor burden
. The technology assists in producing the best patient
outcomes due to the ability to accurately measure the remaining liver tissue that constitutes a considerable
determinant of the surgical planning [17]. The CT perfusion imaging methods can be used to assess the
parameters of blood flow in order to differentiate between benign and malignant liver lesions based on the
measurements of blood flow and blood volume and mean transit time. Quantitative tests are beneficial in not
only knowing about the flow of blood through the lesion but they also aid in the monitoring of therapeutic
response [21].
The MDCT technology has a higher spatial resolution, and an improved temporal resolution compared to the
standard CT scanners that increases its capabilities in detecting small hepatic abnormalities. The system is of
great benefit in the treatment of hepatobiliary and pancreatic diseases by the way of planning before treatment
and checking after the surgeries [17]. Also, the development of the image processing methods, including the use
of anisotropic diffusion filters and morphological operations, has allowed the better identification of malignant
tumors, which allows timely detection and intervention [22]. The CT imaging has a lot of advantages to medical
diagnosis and there are certain performance limitations. Clinical CT imaging applications are hampered by the
use of contrast agents and the limitation of radiation exposure and reduced sensitivity to small lesions.
The concept of CT imaging with contrast enhancement built into CAD models has served to assist the medical
staff in identifying the benign and malign a lesion in the liver more efficiently. This development makes the new
diagnostic framework automation possible [23]. CT is an important diagnostic tool in liver pathology
identification despite the limitations it possesses because its combination with AI-based methods will increase
its diagnostic capabilities to improve patient outcomes and clinical practice.
Magnetic Resonance Imaging
MRI is vital in the detection and characterization of liver pathologies since it assists in the separation of focal
liver lesions such as HCC and other abnormalities of the liver. Its superior imaging technology, particularly when
used with hepatobiliary-specific contrast agents, like gadoxetic acid, not only improves the accuracy of diagnosis
but also provides information on liver diseases in detail. Multiparametric MRI combines several MRI techniques
to enable the use of images of the tumor structure and blood flow, and cellular characteristics to influence the
decision made by healthcare professionals before surgery in relation to HCC patients [24]. The outstanding
quality of MRI in distinguishing between lesions in hepatoma is due to the fact that it is capable of distinguishing
between atypical hemangiomas and metastatic deposits based on the characteristics of the signals presented as
T1-weighted and T2-weighted images. Contrast-enhanced MRI makes patient management more effective
because it indicates the patterns of vascularization which is crucial in the process of diagnosis [25].
Hepatobiliary-specific contrast agents that are administered results in the best lesion detectability as studies show
that MRI has a high sensitivity and specificity level [26]. Diffusion-weighted imaging and magnetic resonance
Page 1358
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
elastography as the types of innovative methods of MRI are combined to help professionals perform detailed
examination of liver lesions based on blood supply and characteristics of tissues.
AMRI is an abbreviated form
of MRI which has demonstrated potential because it is a promising tool that is better in identifying focal lesions
in the liver, as compared to ultrasound use in high-risk groups.
Although MRI is superior to other imaging techniques, it has limitations of cost, availability and motion artifacts,
which restricts its prevalence in the clinical setting. The susceptibility to noise, the necessity to replace basic
imaging methods with more complex ones to improve interpretability are still the fields of research. More so,
MRI-guided interventions are generally more time-consuming than CT-guided procedures, whereas the former
usually offer superior precision, which may be a limitation in emergency clinical practice [48]. The development
of imaging technology and its availability is crucial to maximize the role of MRI in the diagnosis and treatment
of liver disease.
Image Preprocessing
Preprocessing is the essential bridge between raw radiological imaging and model robustness. In CT imaging,
strategies for mitigating beam-hardening artifacts and normalizing pixel intensities are standard for reducing
inter-scanner variability [27]. In the case of MRI, the bias field correction and noise reduction are critical factors
in retaining soft-tissue contrast [27]. Although histogram equalization (HE) and Contrast Limited Adaptive
Histogram Equalization (CLAHE) are frequently used to improve contrast they may cause amplification of noise;
modern architectures often combine the use of spatial filters with either wavelet transforms or Generative
Adversial Networks -based systems to avoid this issue and retain fine anatomical detail [28]. In addition,
standardized normalization, and specifically Z-score scaling, is essential to obtain reproducible fibrosis feature
identifications across multi-center cohorts.
Feature Extraction
A paradigm shift is evident from manual, handcrafted radiomics toward automated feature extraction. Manual
extraction, driven by radiologist-defined Regions of Interest (ROIs), utilizes statistical (first-order), textural
(GLCM-based), and morphological (higher-order) descriptors to capture subtle tissue heterogeneity [29]. They
are particularly applicable in clinical environments with small samples where DL model is vulnerable to
overfitting [16]. On the other hand, automated extraction through CNNs allows hierarchical spatial features to
be obtained without human involvement [30]. Although the sensitivity of DL architectures is stronger in the
context of segmentation, the latter require large scale, annotated datasets to achieve stable performance [31].
Increasingly, research is taking a hybrid approach, using semantic features of radiomics (shape, location) in
conjunction with deep-learned representations, to gain even greater accuracy in classification.
Feature Selection
The final phase of model interpretability and data redundancy minimization depends on feature selection.
Radiomic datasets are high-dimensional, and dimensionality reduction methods, including Principal Component
Analysis (PCA) and Linear Discriminant Analysis (LDA), are customary to simplify model inputs [32]. Sparse
PCA has been especially successful in isolating clinically significant variables without compromising
classification performance [33]. At the same time, LDA is often used to make the most of the separability
between classes, which shows a positive effect on the precision of ML-based classifiers like Random Forests
and SVMs. As more recent empirical studies confirm, intelligent feature selection, even in DL-dominated
architectures, may still be a limitation that plays a critical role in improving diagnostic robustness when
performing diagnosis in a variety of imaging modalities [34].
Feature Extraction
Feature extraction serves as the bridge between raw imaging data and predictive clinical modeling. In liver
pathology, researchers distinguish between semantic features (interpretable radiological characteristics such as
lesion shape, location, and vascularity) and agnostic features (quantitative measures of tissue heterogeneity)
[35]. While semantic analysis remains primarily the domain of expert radiologists, agnostic feature extraction
Page 1359
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
has become the foundation of computational liver diagnostics, enabling the quantification of structural and
textural patterns that are often imperceptible to the human eye.
These agnostic features can be traditionally classified as statistical, transform-based, and structural analysis types
[35]. Manual or automated extraction methodology is also another critical aspect in CAD design. Handcrafted
radiomics Manual feature extraction Manual feature extraction is based on first-order (histogram-based) and
higher-order (e.g., GLCM, GLRLM) statistical descriptors. These techniques have better interpretability and are
especially useful with small and single-center datasets, where deep learning models are prone to overfitting [36].
Conversely, automated feature extraction leverages DL architectures to capture hierarchical, abstract
representations of liver tissue without explicit manual feature engineering. This method is useful in reducing
statistical bias that can be introduced by humans and has proven quite successful in applications like HCC
detection and fibrosis staging [30]. The automated extraction, however, requires massive and heterogeneous
datasets to guarantee the generalizability of the model. The existing evidence indicates that, although the DL-
based automated extraction increasingly gains increased popularity due to its high diagnostic accuracy,
handcrafted radiomic features continue to play a vital role in clinical practice because of their stability, fewer
mathematical demands, and ability to support clinical decision-making. The flow of the radiomic analysis with
the application of the two methods is depicted in Figure 3.
Figure 3: Workflow of radiomics analysis comparing manual and automated feature extraction approaches.
Manual Feature Extraction
Handcrafted radiomics remain highly relevant in liver pathology due to their established diagnostic utility in
identifying structural and intensity-based tissue variations. They are typically grouped into three large categories:
first-order (statistical), second-order (textural), and transform-based, each of which offers a different
understanding of liver disease progression [35]. First-order features are the mean, variance, skewness and
entropy properties of underlying pixel intensity distribution over a specified region of interest (ROI). Although
they are computationally simple, these measures are useful in baseline differentiation of liver tissue states [29].
Second-order features are used to represent more intricate disease signatures by leveraging the spatial
relationships between adjacent pixels, with the most widely used being Gray-Level Co-occurrence Matrices
(GLCM). GLCM features, including contrast, homogeneity, and cluster shade, are important in the process of
measuring subtle irregularities of the texture reflecting fibrosis and steatosis [29]. To perform more advanced
diagnostics, more advanced features (e.g., NGTDM, GLRLM, GLSZM) are used to detect more complex spatial
Page 1360
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
patters in liver tumors and tissue heterogeneity that cannot be revealed by lower-order statistical tools. These are
complemented by transform-based features, which use the Wavelet transforms, Gabor filters, Local Binary
Patterns (LBP) and the Histograms of Oriented Gradients (HOG) to analyze frequency distributions [29]. These
transform-based methods are able to radically improve the definition of the tumor boundaries and the description
of lesions margins by isolating certain spatial frequencies [37]. This combination of complementary feature sets
(summarized in Figure 4) is critical to building strong classification models that demand high-fidelity
descriptions of liver lesions.
Figure 4: Categories of radiomic features commonly utilized for liver disease diagnosis.
Automated Feature Extraction
DL-based automated feature extraction has revolutionized liver CAD systems through its ability to hierarchically
represent imaging data without manually engineered features. Compared to handcrafted radiomics, CNNs are
able to learn hierarchy of spatial features, which has made remarkable progress in segmentation and classification
of liver pathologies across CT, MRI, and ultrasound [30]. Recent research has turned to more advanced CNN
models to enhance diagnostic sensitivity. Architectures like AlexNet, VGGNet, and ResNet have become the
standards in liver pathology detection [38]. The superiority of DL models is supported by empirical evidence;
e.g., Li et al. (2020) show that the sensitivity of HCC classification is 92.16 with Inception-V1 [39]. Moreover,
comparative analyses, including Wang et al. (2018), have indicated that more complex models such as the ResNet
always perform better than the simple models in the characterization of complex focal liver lesions [40]. In order
to overcome a lack of data, transfer learning with pre-trained networks such as GoogLeNet and DenseNet--has
become a typical procedure and has greatly improved the results in poor-data clinical cohorts [30]. Table 1
describes CNN layers. The implementation of CNN-based systems in clinical workflows is limited by three main
challenges, even though they are very accurate:
1. Data Requirements and Overfitting: CNNs with high performance require large, labeled datasets. In
medical imaging where this data is sparse, models are extremely vulnerable to overfitting. Data
augmentation, dropout layers, and strong regularization strategies are compulsory to ensure model
robustness [41].
Page 1361
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
2. Computational Demand: CNN training requires high-performance computing infrastructure, which is a
limitation to practice in a standard hospital environment.
3. The Black-Box Limit: The black-box nature of deep learning decision-making has continued to be a barrier
to clinical trust. Since CNNs are automatically trained on internal patterns, they lack transparency, which
makes it difficult to match AI predictions to standard clinical reasoning [42].
As a result, explainable AI (XAI) and hybrid modeling are becoming more central to the future of liver CAD
research to ensure that the balance between the accuracy of automated diagnostics and the responsibility of
clinicians is achieved.
Table 1- Functions and descriptions of CNN layers [43]
CNN Layer
Description
Convolutional Layer
Fundamental Component
It locates the set of features by scanning the image and performing convolutional
filtering
Pooling Layer
Sub sampling layer usually placed after the convolutional layer.
It receives the feature input from the convolutional layer and applied pooling
operation which works on image size reduction while preserving the important
characteristics of the image.
Activation Layer
demonstrates that the dependent variable as well as positive input have a linear
relationship.
Fully Connected Layer
It is always fully connected and has its rol in classification.
It recieves a vector from different feature maps which are computed by convolutional
and pooling layer.
Feature Selection
Although DL structures automatically extract features, they do not necessarily ensure optimal diagnostic
efficiency. Redundant or irrelevant features in liver CAD modeling are also likely to cause noise, contributing
to the potential overfitting issue due to the relatively small size of most public liver imaging datasets. Therefore,
dimensionality reduction methods, including Principal Component Analysis (PCA) and Linear Discriminant
Analysis (LDA), become necessary to enhance model predictability and interpretability.
PCA has been used as the criterion to identify the most useful texture descriptors in imaging modalities.
Indicatively, research has effectively employed PCA to enhance the distinction of normal liver tissue, chronic
active hepatitis, and cirrhosis [44]. However, the linearity of conventional PCA might not be able to identify
non-linear relationships between features, making it necessary to adopt non-linear methods, such as mutual
information measures or Sparse PCA, to ensure that only the most diagnostically meaningful features are selected
[44].
On the other hand, the LDA is often favored in clinical classification where it explicitly maximizes the separation
between classes which directly improves the predictive power of the chronic liver disease models [45]. LDA can
be much more interpretable than PCA, as it tends to retain information that is specific to a class and reduce
dimensionality. Recent literature emphasises synergistic potential of such techniques e.g. high classification
accuracies (up to 98.31) have been reported when combing LDA with Random Forest models [45]. Moreover,
the most recent studies indicate that a combination of PCA and LDA can create an optimal feature space that
greatly enhances the performance of CNN-based architectures [46].
The successful implementation of these dimensionality reduction methods is not only a computational
optimization but a clinical necessity. Through the elimination of irrelevant data, researchers develop more
generalized and less affected by the black-box constraints of the high-dimensional DL architecture. This
improves the accuracy of the diagnostic output and the ability of these models to be incorporated into a wider
clinical decision support system.
Page 1362
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Classification
The last phase of the CAD pipeline is training classifiers to project extracted features onto a particular clinical
outcome. Many different ML models have been utilized in liver pathology, and their choice has often been
determined by the amount of data, the dimensionality of features, and the need to have clinical interpretability.
Support Vector Machines (SVM) continue to be a base algorithm in high-dimensional, multifaceted data, and
studies have shown up to 90 percent accuracy on steatosis classification [47]. To mitigate the clinical issue of
the imbalance in classes, SVMs are often used with the Synthetic Minority Over-sampling Technique (SMOTE)
that has proven its efficacy in improving predictive stability [48]. Conversely, tree-based models provide a more
transparent method of clinical decision-making. Decision Trees (DT) are valued due to zero-parameter
constraints and have performed better than certain neural network architectures in a few CT-based hepatic lesion
detect tasks [49]. Random Forest (RF) builds on this by leveraging frameworks of ensemble to provide robust
classification; it is especially efficient in the assembly of heterogeneous demographic and laboratory data that
has a consistent rate of accuracy ranging between 72% and 88% [50]. In situations that demand computational
efficiency, Naive Bayes classifiers provide a quick, though sometimes constrained, answer. Its performance is
very much dependent on data characteristics, frequently performing poorly when high feature correlation exists
a familiar situation in medical imaging.
On the other hand, K-Nearest Neighbors (K-NN) is also a non-parametric alternative to SVMs that has proven
superior in a few radiological data. This has been further refined into Advanced variants, like Variable-Neighbor
Weighted Fuzzy K-NN, by adapting weight on the neighbors, but it is again sensitive to the original choice of
distance measures [51]. Recent directions in the field of research prefer ensemble learning and neural architecture
to achieve optimal diagnostic accuracy. Methods such as boosting (e.g., Extreme Gradient Boosting) and bagging
have been found to reduce errors better in predicting liver diseases, combined models have reported an accuracy
of up to 93% [52]. Moreover, Artificial Neural Networks (ANNs) such as CNNs that use spatial data and RNNs
that use sequential data have established new standards in terms of detection accuracy. These neural models are
very good at determining non-linear relationships, but their black-box quality is also a major challenge to clinical
implementation. As a result, classification models are typically selected based on the trade-off between the
predictive capabilities of the neural network at a high level and the transparency and robustness of more
traditional ensemble and tree based methods as summarized in Table 2.
Table 2 Summary of classification models used for liver disease diagnosis.
Evaluation Parameters for liver pathologies diagnosis
The accuracy, recall, confusion matrix, precision, F1 score, and AUC are among the evaluation criteria used to
assess performance of different ML and DL models. One of the easiest measures to understand is accuracy, which
Page 1363
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
offers a broad indication of a model's overall performance by calculating percentage of correct predictions among
all classifications. It reflects how frequently model’s predictions correspond to actual results when considering
both true positives and true negatives. However, accuracy alone may not adequately convey model performance,
particularly when dealing with imbalanced datasets [53]. In such cases, precision becomes important, especially
when false positives carry high costs. Precision is the percentage of correctly predicted positive cases out of all
cases the model classified as positive. The metric holds special value when precise positive prediction accuracy
matters because it focuses on reducing superfluous positive predictions from the model. Recall determines how
well the model identifies positive cases through its assessment of which actual positive cases the model correctly
predicted [54]. The detection of positive cases matters most when failing to identify them would result in serious
outcomes because this measure shows how well the model detects relevant instances. The F1 score combines
precision and recall through a harmonic mean to provide an equilibrium metric between the two metrics. F1
proves highly beneficial for situations with class imbalance because it prevents either precision or recall from
being misrepresented [53].
The confusion matrix enables deep target comparison between actual values and model predictions in
classification model analysis. Actual classes in confusion matrix occupy rows and the predicted classes occupy
columns. The confusion matrix shows how many times a model correctly identifies its predictions while also
showing the count of incorrect predictions known as false positives and negatives and true positives and
negatives. Multiple performance metrics including accuracy, recall, precision and F1 score can be calculated
from the initial values obtained from the model evaluation process [53]. The performance measurement known
as area under the receiver operating characteristic curve (AUC-ROC) determines the likelihood of correctly
placing positive examples higher than negative examples when randomly selected. ROC curve graphically
illustrates how the true-positive rate (sensitivity) changes with the false-positive rate (specificity) at different
threshold settings [54]. A higher AUC indicates a better discriminative ability of the model in distinguishing
between positive and negative classes, thereby summarizing the model’s overall classification performance [54].
These assessment factors are summarized in Table 3, which also describes the applicability and usefulness of
each statistic in evaluating model performance.
Table 3 Performance metrics for evaluating machine learning and deep learning models in liver disease diagnosis
[115]
S.No
Metric
Definition
1.
Accuracy
Accuracy is the proportion of correct predictions out of the total predictions made.
Where:
TP = True Positives
TN=True Negatives
FP=False Positives
FN=False Negative
2.
Precision
Out of all expected positive cases, precision quantifies the percentage of accurately
predicted positive situations.
3.
Recall
Out of all actual positive cases, recall quantifies percentage of accurately predicted
positive cases.
4.
F1-Score
The F1 score has been determined as the precision and recall harmonic means. It offers a
harmony between recall and precision.
Page 1364
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
5.
Area
Under
Curve
The likelihood that a randomly selected positive instance will be ranked higher than a
randomly selected negative instance is represented by the AUC. Plotting the genuine
positive rate (sensitivity) against the false positive rate (specificity) at different threshold
values is what the ROC curve is.
Recent Advances in Machine Learning and Deep Learning for Liver Disease Diagnosis:
As per the review of literature, CAD plays a vital role in supporting the subjective interpretation of radiological
images for liver disease diagnosis. In recent years, the application of image processing algorithms in the medical
field particularly in liver disease detection has become a critical area of focus. ML models such as SVM, RF,
Naïve Bayes and KNN have demonstrated significant potential in classifying liver-related conditions. For
instance, SVM achieved an accuracy of up to 96.27% in identifying liver cirrhosis from multimodal imaging
datasets, and RF reached up to 99.6% accuracy in HCC detection using CT images [55]. Similarly, ensemble
models combining radiomic features and clinical data reported AUCs as high as 0.9975 with XGBoost for HCC
classification. DL models, especially CNNs, have revolutionized the diagnostic process by enabling automatic
feature extraction and high-precision image analysis [56]. Furthermore, transfer learning using architectures such
as ResNet50, MobileNetV2, and DenseNet201 has demonstrated classification accuracies up to 93.75% for
hepatic steatosis stages in NAFLD patients using ultrasound images [57]. The detailed comparative performance
of ML and DL models across different studies is summarized in Tables 4 and 5.
Table 4: Performance of various machine learning models for liver disease classification across different imaging
modalities.
Reference
No./Year of
Publication
Pathology
Feature
Extracted
Classifier
Used
Performance
[58]/2025
Microvascular
Invasion
(MVI) in
HCC
First and
second order
radiomics
features (672
variables),
PCA reduced
to 58
dimensions
RF, MLP
(NeuralNet),
XGB
RF: Acc = 96.8% (Sens:
95.2%, Spec: 97.6%)
XGB: Acc = 68.7% (Sens:
38.1%, Spec: 83.7%)
NeuralNet: Acc = 50%
(Sens: 52.3%, Spec:
48.8%)
[55]/2025
Liver
Cirrhosis
GLCM
ANN, SVM
Accuracy:
ANN: 0.9478
SVM: 0.9627
[59]/2025
HCC
Radiomic
features,
vectorized
tabular data
XGBoost
Accuracy: 0.89±0.05,
AUC: 0.93±0.03
[60]/ 2025
HCC,
Regenerative
Nodules
(RNs),
Dysplastic
Nodules
(DNs)
2264
Radiomic
Features,
Semantic
Features
Machine
Learning
(Five-fold
Cross-
validation)
Combined Model: AUC =
0.896, Radiomics-based
Model: AUC = 0.859,
Semantic Feature-based
Model: AUC = 0.883
[61]/2025
Rectal
Metachronous
MR
Radiomics
Generalized
Linear
GLRM: AUC = 0.765
(Training), AUC = 0.767
Page 1365
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Liver
Metastasis
Regression
Model
(GLRM),
RF
(Validation) RF: AUC =
0.919 (Training), AUC =
0.901 (Validation)
[62]/2025
Fatty Liver
(Mild,
Moderate,
Severe)
2D and 3D
Radiomics
Features
RF, Bagging
DT
RF (2D Radiomics
Model): AUC = 0.973
Bagging DT (2D
Radiomics Model):
Sensitivity = 0.873,
Specificity = 0.939,
Accuracy = 0.864,
Precision = 0.880, F1
Score = 0.876
[63]/2025
Severe
coronary
artery stenosis
in T2DM
patients with
NAFLD
Shape, first-
order
statistics,
texture,
wavelet
features
(radiomics),
and clinical
features (e.g.,
diabetes
duration,
GLS, LDL-
C)
SVM (for
clinical
model),
XGBoost (for
radiomics and
combined
models)
Clinical model AUC:
0.747 (SVM); Radiomics
model AUC: 0.838
(XGBoost); Combined
model AUC: 0.883
(XGBoost); highest F1
score, accuracy, and
precision in validation
[64]/2024
Liver Fibrosis
107
Radiomic
Features
Gradient
Boosted Tree
Model
AUC: 0.997–0.998
(training), 0.617–0.830
(test), Highest AUC: 0.830
(95% CI 0.520–0.830) for
grade 2 fibrosis
classification
[65]/2024
Proliferative
HCC
Original and
Delta
Radiolics
Features
Machine
Learning
Algorithms,
Logistic
Regression
AUC: 0.838 (Training),
0.801 (Validation)
[66]/2024
HCC
Histogram,
Run-length,
Co-
occurance,
wavelet
transform
RF, Boost,
DT, SVM
Accuracy-
(DT: 96.5%
RF: 99.6%
Boost: 99.7%
SVM: 98.0%)
[67]/2024
HCC
GLCM,
GLDM, First
Order
Statistics,
(GLSZM),
GLRLM, and
NGTDM
SVM,
Logistic
regression,
RF, Adaboost,
Xgboost, and
naive Bayes
algorithms
AUC-
Xgboot-0.9975
SVM-0.9825
RF- 0.9861
Logistic Regression-
0.9727
[68]/2023
NAFLD
Ultrasound
image-based
features from
SVM
Sensitivity: 72.2%,
Specificity: 94.6%,
Page 1366
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
10 liver
regions per
subject
Accuracy: 83.4%, PPV:
93.1%, NPV: 77.3%
[69]/2023
HCC
GLCM,
GLRLM,
GLDM,
First-order
statistics,
GLSZM and
NGTDM
RF, XGBoost,
and SVM
AUC MRI-
(XGBoost-0.917, RF-
0.979, SVM-0.961)
Accuracy MRI-
(XGBoost-88%, RF-88%,
SVM-94%)
AUC CT-
(XGBoost-0.822, RF-
0.860, XGBoost-0.938)
Accuracy MRI-
(XGBoost-84%, RF-48%,
SVM-96%)
[70]/ 2023
HCC
833 radiomic
features
RF
AUC: 0.70 ± 0.09
[71]/ 2022
Fibrosis,
Steatosis,
Normal
Computer-
extracted
texture
features
Logistic
Regression,
RF
Logistic Regression (2-
class): AUC = 0.928; RF
(multi-class): AUC =
0.917
[72]/2022
Fatty liver
disease
FOS ,GLCM
Voting Based
Classifier
F1-score 95.64%
Precision-94.28%
Sensitivity-97.05%
Accuracy-95.71%
Specificity-94.44%
[73]/2022
Portal
Hypertension
in Cirrhosis
Patients
FOS and 4
wavelet
features
Logistic
Regression
Accuracy-65.8%
Sensitivity-89.9%
Specificity-33.3%
[74]/2021
Primary liver
cancer vs.
metastatic
liver cancer
First-order,
two-
dimensional
shapeGLCM,
GLRLM,
gray-level
size-zone
matrix,
NGTDM,
GLDM
KNN,
Logistic
Regression
(LR), MLP,
RF, SVM
LR: Accuracy 0.843 ±
0.078, AUC 0.816 ± 0.088,
Sensitivity 0.768 ± 0.232,
Specificity 0.880 ± 0.117
[75]/2019
HCC and liver
abscess
GLCM and
GRLCM
(i) sequential
forward
selection
(SFS), (ii)
sequential
backward
selection
(SBS), and
(iii) F-score)
Accuracy-
(SFS- 89.25%)
(SBS-88.87%)
F-Score- 88.87%
Page 1367
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Table 5: Performance of various deep learning models for liver disease classification based on imaging datasets.
Reference
No. /Year
of Publish
Images
utilised
Pathology
Deep Learning Model
Performance
[76]/2025
MRI
Images
Focal Liver
Diseases
DL Model
Dice coefficient: 0.62; Detection
rate: DL+ Radiologist (0.894) vs.
Radiologist alone (0.825);
Sensitivity: DL (0.883) vs.
Radiologist (0.806); Sensitivity
for lesions <20mm: 0.848 vs.
Radiologists: 0.788; Detection
for lesions ≥20mm: DL (0.867)
vs. Radiologists (0.881), P =
0.671
[77]/2025
CT Images
HCC
3-D Convolutional
Block Attention Module
(CBAM)
Internal validation AUC: 0.807
(95% CI 0.772-0.841),
Radiologist AUC: 0.851 (95%
CI 0.820-0.882), At-risk patient
AUC: 0.769, Indeterminate
scans AUC: 0.815, Small lesions
<2 cm AUC: 0.773, External
testing AUC: 0.789 (95% CI
0.750-0.827)
[78]/2025
CT Images
HCC
Spatio-Temporal 3D
Convolution Network
Internal validation: AUC 0.919
(observation level), 0.901
(patient level); External testing:
AUC 0.901; Compared to
radiological interpretation (AUC
0.839 & 0.822); Negative
predictive values: 0.966
(observation) & 0.951 (patient
level); Observation-level AUCs
for at-risk patients, 2–5 cm
lesions, and singular
portovenous phase: 0.899, 0.872,
and 0.912, respectively.
[79]/2025
CT Images
Liver Tumors
U-Net, Detectron2
U-Net: Mask IoU = 0.903
(effective in simpler cases);
Detectron2: Mask IoU = 0.974
(better in complex cases with
segmented liver regions)
[80]/2025
MRI
Images
MVI in HCC
CNN, TopoCNN,
TopoCNN+Clinic
For tumours ≤ 3.0 cm:
TopoCNN: 0.879 (internal),
0.763 (external)
TopoCNN+Clinic: 0.929
(internal), 0.758 (external)
TopoCNN: 0.890 (internal),
0.871 (external)
TopoCNN+Clinic: 0.895
(internal), 0.879 (external)
Page 1368
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
[81]/2025
CT Images
Fatty Liver
2D & 3D DL models,
2D & 3D radiomics
models
Best AUC: 0.973 (2D radiomics
model with random forest)
Best Sensitivity: 0.873
Best Specificity: 0.939
Best Accuracy: 0.864
Best Precision: 0.880
Best F1 Score: 0.876 (2D
radiomics model with Bagging
decision tree)
[82]/2025
MRI
Images
HCC
Swin Transformer
AUC: 0.77-0.79 (Radiomics),
AUC: 0.79 (Pathomics), C
index: 0.69 (Training), 0.60
(Internal), 0.67 (External), Time-
dependent AUCs for 3-year PFS:
0.83, 0.81, 0.78
[83]/2025
CT Images
Focal Liver Lesions
GAN for data
augmentation
- Localization: Mean Average
Precision = 0.81
- Multiclass Classification
Accuracy: 0.97 (95% CI: 0.95-
0.99)
- Accuracy for FLLs ≤ 3 cm: 0.83
(95% CI: 0.68-0.98)
- Accuracy for FLLs > 3 cm: 0.87
(95% CI: 0.77-0.97)
- Classification Accuracy for
FLLs ≤ 3 cm: 0.95 (95% CI:
0.92-0.98)
- Classification Accuracy for
FLLs > 3 cm: 0.97 (95% CI:
0.94-1.00)
[66]/2025
Ultrasound
Images
Hepatic Steatosis
(HS) in NAFLD
patients
InceptionV3,
MobileNetV2,
ResNet50,
DenseNet201,
NASNetMobile
- 89.15%-93.75% (S0-S1 vs. S2-
S3) with augmentation
- 79.69%-91.21% (S0 vs. S1 vs.
S2 vs. S3) with augmentation
- 80.45%-82.73% (S0-S1 vs. S2-
S3) without augmentation
- 59.54%-63.64% (S0 vs. S1 vs.
S2 vs. S3) without augmentation
- HRI measurement by
radiologists: 82% (S ≥ S1),
91.56% (S ≥ S2), 96.19% (S =
S3)
[84]/2024
CT and
MRI
Images
HCC – MVI
prediction
Multimodal DL using
DenseNet121 + ELM;
Transfer Learning
applied
AUC = 0.844 (MDL model);
AUC = 0.871 (MDL + clinical
features); Outperformed single-
modality models (AUC = 0.706–
0.776 CT, 0.706–0.717 MRI)
[85]/2024
MRI
Images
HCC
DL Model
Improved image quality and
detection rates (91.4%-93.4%)
compared to PI-DWI and CS-
DWI; especially effective in
hepatic dome lesions (94.8%-
97.4%)
Page 1369
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
[86]/2024
CT Images
HCC, intrahepatic
cholangiocarcinoma
(ICC), metastatic
tumors (MET),
focal nodular
hyperplasia (FNH),
hemangioma
(HEM), cysts
(CYST)
Liver Lesion Network
(LiLNet)
Benign vs Malignant: ACC =
94.7%, AUC = 97.2%
HCC, ICC, MET: ACC = 88.7%,
AUC = 95.6%
FNH, HEM, CYST: ACC =
88.6%, AUC = 95.9%
[87]/2023
CT and
MRI
Images
MVI in HCC
ResNet18
(DLCT_ALL,
DLMRI_ALL, DLCT +
MRI), SVM (CALL)
DLCT + MRI: AUC 0.819 (vs.
DLCT_ALL 0.742);
DLMRI_ALL: AUC 0.794 (vs.
RMRI 0.766); CALL model
significant for prognosis (P <
0.001)
[88]/2023
MRI
Images
Early-stage HCC
recurrence
VGG16 + XGBoost
AUC-ROC: 0.71–0.85; 5/6
models significant for RFS (p <
0.05)
[89]/ 2022
MRI
Images
Focal liver lesions
EfficientNetB0
AUC: 0.84 (±0.1), Sensitivity:
0.78 (±0.14), Specificity: 0.86
(±0.08), NPV: 0.89 (±0.08),
PPV: 0.71 (±0.17)
[90]/2022
CT Images
MVI,HCC
Pretrained CNNs via
transfer learning; SVM
classifier
MVI Prediction: AUC 0.909,
Accuracy 96.47%, Sensitivity
90.91%, Specificity 97.30%,
PPV 83.33%, NPV 98.63%
[91]/2022
CT Images
Colorectal Liver
Metastasis
(CRLM), Ablation
Zone
Hybrid-WNet (also
compared: 3D-UNet,
Residual 3D-UNet,
Dense 3D-UNet)
DSC (3D-IRCADb): 0.73 (0.41–
0.88), Global DSC (LiTS):
0.810, Surface distance: 1.75
mm, Sensitivity for lesions ≥15
mm: 98%, Likert score ≥4 for
CRLM: 100%, ablation zones:
84%
[92]/2021
CT Images
Small HCC (sHCC
≤2 cm) in cirrhotic
liver
PM-DL (Pattern
Matching + Deep
Learning using CNN)
Sensitivity: 89.74%, PPV:
85.00%, DICE coefficient: 0.77
± 0.16.
[93]/2021
CT Images
Microvascular
Invasion (MVI) in
HCC
3D-CNN, XGBoost
Training AUROC: 3D-CNN –
0.980; RRC (XGBoost) – 0.952
Validation AUROC: 3D-CNN –
0.906; RRC – 0.887
[94]/2020
CT and
MRI
Images
Clinically
Significant Portal
Hypertension
(CSPH)
Deep CNN
AUC: 0.998 (Train), 0.912
(Validation), 0.933 (Test)
Critical Comparison between Machine Learning and Deep Learning models.
Although the literature reviewed in this work is rapidly growing, with a significant number of studies appearing
in Tables 4 and 5, a critical review indicates that the popular discourse of the superiority of deep learning (DL)
in the diagnosis of liver diseases is not always backed up by empirical evidence. Rather, model performance
highly depends on the size of datasets, imaging modality, strategy feature engineering, and rigor of validation-
something that is not adequately recognized. Although, unlike other image-based models, DL models,
specifically CNN-based models, prove to be highly beneficial in lesion localization, lesion segmentation, and
detection of tiny or hidden abnormalities, they are not universally more beneficial than other image-based models
Page 1370
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
in terms of diagnostic goals. Radiomics-driven machine learning (ML) pipelines are also able to match or surpass
end-to-end DL pipelines in several CT-based and MRI-based studies, especially in classification, including HCC
detection, fibrosis staging, and MVI prediction. Interestingly, RF and XGBoost models are repeated to show
AUC scores above 0.95, sometimes nearly 0.99, whilst numerous DL models level off at similar or worse
performance after being trained on small or single centre datasets. The observation contradicts the belief on
which the superior performance of a diagnostic can be ensured solely on architectural complexity.
The enduring effectiveness of ML models highlights the ongoing usefulness of handcrafted radiomic features,
particularly in clinical environments, where dataset sizes are limited by their nature. ML techniques based on
radiomics have the advantages of: lower data requirements, greater training stability, and enhanced inter-cohort
reproducibility.
By contrast, DL models often use aggressive data augmentation or transfer learning in order to cover inadequate
training samples, which brings up issues of model overfitting and domain shift. In several real-life liver imaging
cases, therefore, ML models provide a more realistic performance-to-feasibility trade-off.
A very evident interaction of modalities-model interaction is also delivered in the synthesis but is mostly
neglected in individual studies. CT-based datasets are always the strongest with the two approaches- ML and
DL, especially with HCC-related tasks. MRI-based DL models are only better at multiparametric and prognostic
but more variable across external validation cohorts. dl models based on ultrasound, although promising, are
highly sensitive to the variability in acquisition as well as operator dependence, and as such are not robust. These
results imply that the choice of models must be made in a modality-conscious manner as opposed to the one-
size-fits-all DL paradigm.
Hybrid ML-dl models often improve standalone models, and the reported AUC improvements are between 3-
8% , which indicates that the feature fusion strategies are more effective than depth-only network improvement.
This underscores a changes in method of models to data driven optimization. The existing evidence lacks
universal dominance in the application of deep learning models over machine learning methods in the diagnosis
of liver diseases. Rather, performance optimization can be achieved through modality-sensitive model selection,
feature engineering based on radiomics and hybrid solutions that combine clinical and imaging information.
Limitations and Challenges of CAD Systems for Liver Pathology Detection
Although there are great advances, CAD liver pathology detection systems have a range of issues that are critical
in preventing their use in clinical settings. One significant shortcoming is that not all training data sets are diverse
due to the lack of diverse populations in the development of many CAD systems resulting in biases and
differences in model performance over other patient demographics, disease prevalence, and imaging modalities.
To mitigate this, it is necessary to implement multi-center datasets collection to enhance the robustness and
external validity of AI models. Moreover, the complexity of deep learning models poses interpretability issues,
where it is necessary to understand how a model has made a decision in order to obtain clinician trust; a variety
of Explainable AI methods such as Grad-CAM, SHAP, and LIME provide some answers, but typically do not
achieve complete transparency, resulting in hesitation to adopt them because they do not align with conventional
clinical reasoning. The regulatory acceptance of AI-based CAD systems is also a demanding process that requires
robust evidence of clinical safety and efficiency, and ethical concerns including patient privacy, data security,
and informed consent should be addressed with caution, especially in a case of multi-institutional data sharing
to conduct model training. Lastly, the implementation of CAD systems into the current Picture Archiving and
Communication Systems (PACS) and hospital processes is practically challenging, with current systems
potentially not fitting AI-powered systems, and active work of CAD developers, medical practitioners, and
regulatory agencies is required to harmonize the standards of a seamless integration of AI-based tools.
Future Perspectives
The application of modern CAD methods in screening liver pathology is changing the face of medical imaging
by boosting the quality and speed of diagnosis. Multimodal imaging integrates CT, MRI and lab findings making
Page 1371
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
it possible to approach the diagnostic process in a holistic manner. The inclusion of multiple data sources makes
CAD models capable of obtaining a more detailed insight of the conditions of the patient and enhances the
accuracy of the diagnosis and clinical decision-making process. Recent models such as contextual convolutional
neural network (COCOn) apply contextual information to more effectively segment liver pathology, surpassing
conventional single-modality models with complementary imaging data. Applications of AI are currently being
spread across different imaging modalities, such as CT and MRI, and enable the identification of focal lesions
and chronic liver disease. The introduction of AI into clinical practice has proven to be able to expand the
diagnostic abilities of radiologists, resulting in better patient outcomes. Federated learning is especially effective
in responding to data sharing challenges when sensitive patient data are involved. The methodology facilitates
the cooperation of various medical facilities and ensures patient privacy to create powerful AI models that can
be implemented on a large scale. XAI is one of the most important aspects that are more significant in this
scenario. To become open in a way that clinicians will trust them and regulators will approve them, AI models
will require XAI that will accelerate their implementation in medical settings. The comparison between ML and
DL models reveals that the patterns of performance of the two models are significant in the detection of liver
disease despite all studies lacking confidence intervals (CIs) or p-values to determine the reliability of
performance. Statistical validation tests such as t-tests and ANOVA should be applied in the research to
understand whether the performance of the DL models is significant at a significant p value of 0.05 or the
difference in the performance is due to the variation in the data sets. Federated learning and self-supervised
methods also enhance CAD applications by resolving issues of data scarcity and privacy issues. Self-supervised
learning allows models to utilize extensive volumes of unlabeled medical imaging data, avoiding the reliance on
costly labeled datasets. Federated learning enables institutions to collaborate but at the same time keep patient
data confidential, which makes AI models more generalizable across diverse populations. Such methods are not
only effective in enhancing the efficiency of CAD models, but also in large-scale implementation in clinical
environments. AI-based personalized medicine is transforming liver disease treatment by forecasting disease
progression and responses to treatment. Machine learning models process data unique to patients (e.g., genetic
profiles, medical history) to personalize interventions, with a maximum accuracy of 80% in identifying liver
disease, greatly outperforming other diagnostic methods [95]. Predictive models can be implemented to ensure
high-risk patients can be identified early, and prompt actions can be taken to improve their prognosis and lower
healthcare expenses. The development of DL architecture, such as Vision Transformers (ViTs), CNN-RNN
hybrid frameworks, and attention-based AI, have further optimized medical imaging analysis, allowing them to
extract more features [96]. These architectures are highly effective in modeling long-range dependencies, as well
as capturing contextual information, and are thus appropriate in the challenging imaging problems in liver
pathology detection. CNNs and RNNs hybrid models enhance the sequential imaging data analysis, thereby
enhancing an understanding of the disease progression [97]. In spite of these developments, there are still
challenges. The need of a large and diverse dataset is to guarantee model generalizability in the context of various
patient groups and imaging modalities. To make AI-driven CAD models useful in the detection of liver
pathology in real-world scenarios, future investigations must center on multi-center clinical trials, standard
evaluation benchmarks, and better techniques of AI interpretability.
CONCLUSION
Incorporation of CAD systems has transformed liver disease diagnosis through enhanced accuracy of diagnosis,
minimized bias in observers and facilitated early intervention. The review identified AI-driven CAD models in
three significant facets of image preprocessing and feature extraction and classification processes. Studies on
ML and DL models have discovered their capability to enhance the diagnosis of non-invasive liver diseases
using CNN-based architectures that yielded high scores in the segmentation and classification tasks. There are
several critical challenges that continue to exist in the wake of these developments. The generalizability of CAD
models is compromised by the lack of variety in the dataset and the absence of multi-centre verification. Also,
black-box characteristics of deep learning models create an important obstacle to clinical acceptance because
clinicians need interpretable and transparent decisions. The regulatory restrictions and complexity of adapting
AI-driven CAD to hospital workflows are also a hindrance to real-life implementation. Research directions
should involve future studies to make models more robust and clinically applicable with the use of federated
learning, multimodal data fusion, and self-supervised learning techniques. A radiological imaging combination
with clinical and molecular biomarkers has the potential to create more specific, explicable, and patient-centric
Page 1372
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
diagnostic models. Radiologists, AI researchers, and policymakers are required to collaborate in order to close
the gap between AI research and clinical practice. Data standardization, interpretability, and real-world validation
will be essential in pushing CAD systems to routine clinical application. As the field of AI-based CAD advances
and becomes more refined, it has the potential to revolutionize the process of managing liver disease and achieve
better patient outcomes and precision medicine in hepatology.
List of abbreviations
Abbreviation
Full Form
HCC
Hepatocellular Carcinoma
ML
Machine Learning
DL
Deep Learning
CAD
Computer-Aided Diagnosis
CNN
Convolutional Neural Networks
AUC
Area Under Curve
AI
Artificial Intelligence
MASLD
Metabolic Dysfunction Associated Steatotic Liver
Disease
NAFLD
Non Alcoholic Fatty Liver Disease
US
Ultrasound
CT
Computed Tomography
MRI
Magnetic Resonance Imaging
MDCT
Multi Detector Computed Tomography
AMRI
Abbreviated Magnetic Resonance Imaging
PCA
Principal Component Analysis
HE
Histogram Equalization
CLAHE
Contrast Limited Adaptive Histogram Equalization
GLCM
Gray-Level Co-occurrence Matrices
NGTDM
Neighborhood Gray-Tone Difference Matrix
GLRLM
Gray-Level Run-Length Matrix
GLSZM
Gray-Level Size Zone Matrix
GLDM
Gray-Level Distance Zone Matrix
LBP
Local Binary Patterns
HOG
Histogram of Oriented Gradients
LDA
Linear Discriminant Analysis
KNN
K-Nearest Neighbor
SVM
Support Vector Machine
DT
Decision Trees
MLP
Multilayer Perceptrons
RF
Random Forest
RNN
Recurrent Neural Networks
ANNs
Artificial Neural Networks
TP
True Positives
TN
True Negatives
FP
False Positives
FN
False Negative
ROC
Receiver Operating Characteristic
ELM
Extreme Learning Machine
PPV
Positive Predictive Value
NPV
Negative Predictive Value
XAI
Explainable Artificial Intelligence
Page 1373
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
ACKNOWLEDGEMENTS
The authors acknowledge that this research did not receive financial or institutional support beyond regular
academic resources.
Author Contribution Statement
SB: Conceptualization, Writing – original draft. SA: Review and editing, Supervision, AS: Review, Editing,
Conceptualization, AP: Conceptualization.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-
profit sectors.
Data Availability Statement
No datasets were generated or analysed during the current study.
Conflicts of Interest
The authors declare no conflicts of interest related to this work.
Declarations
Ethics Approval and Consent to participate
Not applicable.
Consent for publication
Not applicable
Competing interest
The authors declare no competing interests
Author details
1 Chitkara School of Health Sciences, Chitkara University, Punjab, India.
2 Chitkara University Institute of Engineering and Technology,Chitkara University, Punjab, India.
3 Department of Radiology, Fortis Hospital, Mohali
REFERENCES
1. Kalejahi BK, Meshgini S, Danishvar S, Khorram S. Diagnosis of liver disease by computer-assisted
imaging techniques: A literature review. Intelligent Data Analysis. 2022 Jan 1;26(4):1097-114.
2. Devarbhavi H, Asrani SK, Arab JP, Nartey YA, Pose E, Kamath PS. Global burden of liver disease:
2023 update. Journal of hepatology. 2023 Aug 1;79(2):516-37.
3. Asrani SK, Devarbhavi H, Eaton J, Kamath PS. Burden of liver diseases in the world. Journal of
hepatology. 2019 Jan 1;70(1):151-71.
4. Baviskar K, Kshirsagar A, Raut H, Shaikh MRN. Overview: Global burden of liver disease. Int J Pharm
Chem Anal. 2024;11(1):1-10.
5. Tsoneva DK, Ivanov MN, Vinciguerra M. Liquid liver biopsy for disease diagnosis and prognosis.
Journal of Clinical and Translational Hepatology. 2023 Jul 27;11(7):1520.
Page 1374
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
6. Jha M, Mishra HS, Chaudhary AK, Jana D (2020) Diagnosis of liver pathology by ultrasound at dmch,
laheriasarai, bihar. International journal of scientific research 47–48.
7. Chaurasia R (2024) Radiological approach for early diagnosis of liver parenchymal disease. Asian
journal of medical sciences 15:181–186
8. Costa G, Lanza E, Donadon M, Torzilli G (2023) Imaging. Oxford University Press, pp 49-C6P66
9. Maino C, Vernuccio F, Cannella R, et al (2024) Radiomics and liver: Where we are and where we are
headed? European Journal of Radiology 171:111297
10. Zamanian H, Shalbaf A. Grading the Dominant Pathological Indices in Liver Diseases from
Pathological Images Using Radiomics Methods. Frontiers in Biomedical Technologies. 2024 Apr 15.
11. Kumar EP, Rahaman H, Reddy CV, Sahil R, Buchanalli S, Depuru BK (2024) Deep Learning-Based
Liver Histopathology Image Classification: State-of-the-Art Techniques and Emerging Trends.
International journal of innovative science and research technology 1964–1971
12. Nakatsuka T, Tateishi R, Sato M, Hashizume N, Kamada A, Nakano H, Kabeya Y, Yonezawa S, Irie R,
Tsujikawa H, Sumida Y. Deep learning and digital pathology powers prediction of HCC development
in steatotic liver disease. Hepatology. 2025 Mar 1;81(3):976-89.
13. Grignaffini F, Barbuto F, Troiano M, Piazzo L, Simeoni P, Mangini F, De Stefanis C, Onetti Muda A,
Frezza F, Alisi A. The use of artificial intelligence in the liver histopathology field: a systematic review.
Diagnostics. 2024 Feb 10;14(4):388.
14. Doi K. Computer-aided diagnosis in medical imaging: historical review, current status and future
potential. Computerized medical imaging and graphics. 2007 Jun 1;31(4-5):198-211.
15. Shiraishi J, Li Q, Appelbaum D, Doi K. Computer-aided diagnosis and artificial intelligence in clinical
imaging. InSeminars in nuclear medicine 2011 Nov 1 (Vol. 41, No. 6, pp. 449-462).
16. Lowry SB, Joseph S, Psoter KJ, Dunn E, Mansoor S, Smith SK, Karnsakul W, Naguib G, Ng K,
Scheimann AO. Efficacy of Ultrasound for the Detection of Possible Fatty Liver Disease in Children.
Diagnostics. 2024 Jul 31;14(15):1652.
17. Biris AI, Karamatzanis I, Biri D, Biris IA, Maravegias N, Maravegias Sr N. Non-Invasive Ultrasound
Diagnostic Techniques for Steatotic Liver Disease and Focal Liver Lesions: 2D, Colour Doppler, 3D,
Two-Dimensional Shear Wave Elastography (2D-SWE), and Ultrasound-Guided Attenuation
Parameter (UGAP). Cureus. 2024 Oct 22;16(10).
18. Von Ramm OT, Smith SW. Prospects and limitations of diagnostic ultrasound. InRecent and Future
Developments in Medical Imaging II 1979 Dec 26 (Vol. 206, pp. 6-18). SPIE.
19. Peltec A, Sporea I. Multiparametric ultrasound as a new concept of assessment of liver tissue damage.
World Journal of Gastroenterology. 2024 Mar 28;30(12):1663.
20. Lupescu IG, Grasu MC. CT in Hepato-Bilio-Pancreatic Surgical Pathology. InThe IASGO Textbook
of Multi-Disciplinary Management of Hepato-Pancreato-Biliary Diseases 2022 Jun 4 (pp. 99-111).
Singapore: Springer Nature Singapore.
21. Gadupudi V, Ramachandran R, Mohanarangam VS. The Role of Computed Tomography Perfusion in
Various Focal Liver Lesions. Cureus. 2022 Dec 11;14(12).
22. Kumar KP, Prasad KS, Sunil N, Nikhil A. Liver Cancer Detection Using Image Processing.
International Journal For Science Technology And Engineering 2024;12:1139–45.
23. Krishan A, Mittal D. Ensembled liver cancer detection and classification using CT images. Proceedings
of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine. 2021
Feb;235(2):232-44.
24. Jahn M, Layer G. Multiparametric magnetic resonance imaging for hepatocellular carcinoma, part 1:
Morphology and dynamic perfusion imaging in primary diagnostics and treatment monitoring.
Radiologie (Heidelberg, Germany). 2024 Mar 19.
25. Kavaja F, Veselaj F. Exploring the Role of MRI in the Detection of Atypical Liver Hemangiomas and
Exclusion of Metastases. International Journal of Biomedicine 2023;13:165–8.
26. Ichikawa S, Goshima S. Clinical significance of liver MR imaging. Magnetic Resonance in Medical
Sciences. 2023;22(2):157-75.
27. Al Ahad A, Das B, Khan MR, Saha N, Zahid A, Ahmad M. Multiclass liver disease prediction with
adaptive data preprocessing and ensemble modeling. Results in Engineering. 2024 Jun 1;22:102059.
28. Khan RA, Luo Y, Wu FX. Machine learning based liver disease diagnosis: A systematic review.
Neurocomputing. 2022 Jan 11;468:492-509.
Page 1375
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
29. Cai JH, He Y, Zhong XL, Lei H, Wang F, Luo GH, Zhao H, Liu JC. Magnetic resonance texture analysis
in Alzheimer's disease. Academic radiology. 2020 Dec 1;27(12):1774-83.
30. R., Manjunath., Anshul, Ghanshala., Karibasappa, Kwadiki. Deep learning algorithm performance
evaluation in detection and classification of liver disease using CT images. Multimedia Tools and
Applications, (2023). doi: 10.1007/s11042-023-15627-z
31. Sunarjo MS, Gan H, Setiadi DRIM. High-Performance Convolutional Neural Network Model to
Identify COVID-19 in Medical Images. Epub ahead of print 30 August 2023. DOI:
10.33633/jcta.v1i1.8936.
32. Karule PT, Dudul SV. PCA NN Based Classifier For Liver Diseases from Ultrasonic Liver
Images. International Conference on Emerging Trends in Engineering and Technology 2009; 76–80.
33. Cui M, Ma X, Li Q, Hou T, Wang Y. A Novel Principal Component Selection Strategy for Feature
Extraction and Dimension Reduction. Cybersecurity and Cyberforensics Conference. 2022;4040–5.
34. Dougherty G. Feature Extraction and Selection. In Springer, New York, NY; 2013. p. 123–41.
35. Corrias G, Micheletti G, Barberini L, Suri JS, Saba L. Texture analysis imaging “what a clinical
radiologist needs to know”. European Journal of Radiology. 2022 Jan 1;146:110055. 32.
36. Seva A, Rao ST, Sireesha M. Prediction of liver disease with random forest classifier through SMOTE-
ENN balancing. In2024 IEEE 13th International Conference on Communication Systems and Network
Technologies (CSNT) 2024 Apr 6 (pp. 928-933). IEEE.
37. Sorensen L, Shaker SB, De Bruijne M. Quantitative analysis of pulmonary emphysema using local
binary patterns. IEEE transactions on medical imaging. 2010 Feb 2;29(2):559-69.
38. Wu Z. Application of CNN Classic Model in Modern Image Processing. Journal of Advances in
Engineering and Technology 2024;1:1–6.
39. Li J, Wu Y, Shen N, Zhang J, Chen E, Sun J, Deng Z, Zhang Y. A fully automatic computer-aided
diagnosis system for hepatocellular carcinoma using convolutional neural networks. Biocybernetics
and Biomedical Engineering. 2020 Jan 1;40(1):238-48.
40. Wang W, Iwamoto Y, Han X, Chen YW, Chen Q, Liang D, Lin L, Hu H, Zhang Q. Classification of
focal liver lesions using deep learning with fine-tuning. InProceedings of the 2018 International
Conference on digital medicine and image processing 2018 Nov 12 (pp. 56-60).
41. Sunarjo MS, Gan H, Setiadi DRIM. High-Performance Convolutional Neural Network Model to
Identify COVID-19 in Medical Images. Epub ahead of print 30 August 2023. DOI:
10.33633/jcta.v1i1.8936.
42. Patil A, Patil M. A Comprehensive Review on Explainable AI Techniques, Challenges, and Future
Scope. Springer International Publishing, 2023, pp. 517–529.
43. Das AK, Ahmed SS. Convolutional neural networks. 2023. page 186–91.
44. Karule PT, Dudul SV. PCA NN Based Classifier For Liver Diseases from Ultrasonic Liver
Images. International Conference on Emerging Trends in Engineering and Technology 2009; 76–80.
45. Karna A, Khan NZ, Rauniyar R, et al. Unified dimensionality reduction techniques in chronic liver
disease detection. Epub ahead of print 30 December 2024. DOI: 10.48550/arxiv.2412.21156.
46. Dougherty G. Feature Extraction and Selection. In Springer, New York, NY; 2013. p. 123–41.
47. Yeh, W.C.; Jeng, Y.M.; Li, C.H.; Lee, P.H.; Li, P.C. Liver steatosis classification using high-frequency
ultrasound. Ultrasound Med. Biol. 2005, 31, 599–605
48. Nugraha MA, Mazdadi MI, Farmadi A, Saragih TH. SMOTE Class Balance and Selection of Ensemble
Filter Features on Support Vector Machine for Liver Disease Classification. Journal of Information
Technology and Computer Science. 2023 Dec 30; 10(6):1273-84.
49. Das A, Das P, Panda SS, Sabut S. Detection of liver cancer using modified fuzzy clustering and decision
tree classifier in CT images. Pattern Recognition and Image Analysis. 2019 Apr;29:201-11.
50. Wei H. A Random Forest-based Prediction for Liver Cirrhosis. Applied and Computational Engineering
2024;78:98–106.
51. Kumar P, Thakur R. Liver disorder detection using variable- neighbor weighted fuzzy K nearest
neighbor approach. Multimedia Tools and Applications 2021; 80: 16515–16535.
52. Sumellika T, Prasad R. A Combined Ensemble Model (CEM) for a Liver Cancer Detection
System. International Journal of Advanced Computer Science and Applications; 15. Epub ahead of
print 1 January 2024. DOI: 10.14569/ijacsa.2024.0150218.
Page 1376
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
53. Guetari R, Ayari H, Sakly H. Computer-aided diagnosis systems: a comparative study of classical
machine learning versus deep learning-based approaches. Knowledge and Information Systems. 2023
Oct;65(10):3881-921.
54. Chahar S, Roy PK. Covid-19: A comprehensive review of learning models. Archives of Computational
Methods in Engineering. 2022 May;29(3):1915-40.
55. Prabazhini K, Jayakody JA, Sivatharshan M. Predicting the Severity of Liver Cirrhosis with Image
Processing Based Machine Learning Models.
56. Budai BK, Tóth A, Borsos P, Frank VG, Shariati S, Fejér B, Folhoffer A, Szalay F, Bérczi V, Kaposi
PN. Three-dimensional CT texture analysis of anatomic liver segments can differentiate between low-
grade and high-grade fibrosis. BMC Medical Imaging. 2020 Dec;20:1-1.
57. Drazinos P, Gatos I, Katsakiori PF, Tsantis S, Syrmas E, Spiliopoulos S, Karnabatidis D, Theotokas I,
Zoumpoulis P, Hazle JD, Kagadis GC. Comparison of deep learning schemes in grading non-alcoholic
fatty liver disease using B-mode ultrasound hepatorenal window images with liver biopsy as the gold
standard. Physica Medica. 2025 Jan 1;129:104862.
58. Famularo S, Penzo C, Maino C, Milana F, Oliva R, Marescaux J, Diana M, Romano F, Giuliante F,
Ardito F, Grazi GL. Preoperative detection of hepatocellular carcinoma's microvascular invasion on
CT-scan by machine learning and radiomics: A preliminary analysis. European Journal of Surgical
Oncology. 2025 Jan 1;51(1):108274.
59. Wang B, Rodriguez I, Breitinger L, Tollens F, Itzel T, Grimm D, Sirazitdinov A, Frölich M, Schönberg
S, Teufel A, Hesser J. A baseline for machine-learning-based hepatocellular carcinoma diagnosis using
multi-modal clinical data. arXiv preprint arXiv:2501.11535. 2025 Jan 20.
60. Wen X, Pan W, Liu T, Yu C, Sun M, Bian J, Yu C, Luo J. MRI-based radiomics model for predicting
the nature of nodules in cirrhotic liver. Asian Journal of Surgery. 2025 Feb 5.
61. Long ZD, Yu X, Xing ZX, Wang R. Multiparameter magnetic resonance imaging-based radiomics
model for the prediction of rectal cancer metachronous liver metastasis. World Journal of
Gastrointestinal Oncology. 2025 Jan 15;17(1):96598.
62. Zhang H, Liu J, Su D, Bai Z, Wu Y, Ma Y, Miao Q, Wang M, Yang X. Diagnostic of fatty liver using
radiomics and deep learning models on non-contrast abdominal CT. PloS one. 2025 Feb
13;20(2):e0310938.
63. Liang M, Fang L, Chen X, Huang W. Detecting severe coronary artery stenosis in T2DM patients with
NAFLD using cardiac fat radiomics-based machine learning. Scientific Reports. 2025 Feb
25;15(1):6788.
64. Gotta J, Gruenewald LD, Reschke P, Booz C, Mahmoudi S, Stieltjes B, Choi MH, D'Angelo T, Bernatz
S, Vogl TJ, Sinkus R. Noninvasive Grading of Liver Fibrosis Based on Texture Analysis From MRI‐
Derived Radiomics. NMR in Biomedicine. 2025 Jan;38(1):e5301.
65. Yan Z, Liu Z, Zhu G, Lu M, Zhang J, Liu M, Jiang J, Gu C, Wu X, Zhang T, Zhang X. Gadoxetic Acid-
Enhanced MRI-Based Radiomic Models for Preoperative Risk Prediction and Prognostic Assessment
of Proliferative Hepatocellular Carcinoma. Academic Radiology. 2025 Jan 1;32(1):157-69.
66. Malik MH, Ghous H, Rashid T, Maryum B, Hao Z, Umer Q. Feature extraction-based liver tumor
classification using Machine Learning and Deep Learning methods of computed tomography images.
Cogent Engineering. 2024 Dec 31;11(1):2338994.
67. Budai BK. THE ROLE OF CT TEXTURE ANALYSIS IN THE DIAGNOSTICS OF BENIGN AND
MALIGNANT LESIONS OF ABDOMINAL ORGANS (Doctoral dissertation).
68. Tahmasebi A, Wang S, Wessner CE, Vu T, Liu JB, Forsberg F, Civan J, Guglielmo FF, Eisenbrey JR.
Ultrasound‐based machine learning approach for detection of nonalcoholic fatty liver disease. Journal
of Ultrasound in Medicine. 2023 Aug;42(8):1747-56.
69. Zhu L, Wang F, Chen X, Dong Q, Xia N, Chen J, Li Z, Zhu C. Machine learning-based radiomics
analysis of preoperative functional liver reserve with MRI and CT image. BMC medical imaging. 2023
Jul 17;23(1):94.
70. Han YE, Cho Y, Kim MJ, Park BJ, Sung DJ, Han NY, Sim KC, Park YS, Park BN. Hepatocellular
carcinoma pathologic grade prediction using radiomics and machine learning models of gadoxetic acid-
enhanced MRI: a two-center study. Abdominal Radiology. 2023 Jan;48(1):244-56.
71. Sultan LR, Cary TW, Al-Hasani M, Karmacharya MB, Venkatesh SS, Assenmacher CA, Radaelli E,
Sehgal CM. Can Sequential Images from the Same Object Be Used for Training Machine Learning
Page 1377
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
Models? A Case Study for Detecting Liver Disease by Ultrasound Radiomics. AI. 2022 Sep 1;3(3):739-
50
72. Gaber A, Youness HA, Hamdy A, Abdelaal HM, Hassan AM. Automatic classification of fatty liver
disease based on supervised learning and genetic algorithm. Applied Sciences. 2022 Jan 5;12(1):521.
73. Wan S, Wei Y, Zhang X, Yang C, Hu F, Song B. Computed tomography-based texture features for the
risk stratification of portal hypertension and prediction of survival in patients with cirrhosis: a
preliminary study. Frontiers in Medicine. 2022 Apr 1;9:863596.
74. Mao B, Ma J, Duan S, Xia Y, Tao Y, Zhang L. Preoperative classification of primary and metastatic
liver cancer via machine learning-based ultrasound radiomics. European radiology. 2021 Jul;31:4576-
86.
75. Xu SS, Chang CC, Su CT, Phu PQ. Classification of liver diseases based on ultrasound image texture
features. Applied Sciences. 2019 Jan 19;9(2):342.
76. Dai H, Xiao Y, Fu C, Grimm R, von Busch H, Stieltjes B, Choi MH, Xu Z, Chabin G, Yang C, Zeng
M. Deep Learning–Based Approach for Identifying and Measuring Focal Liver Lesions on Contrast‐
Enhanced MRI. Journal of Magnetic Resonance Imaging. 2025 Jan;61(1):111-20.
77. Peng C, Yu PL, Lu J, Cheng HM, Shen XP, Chiu KW, Seto WK. Opportunistic detection of
hepatocellular carcinoma using noncontrast CT and deep learning artificial intelligence. Journal of the
American College of Radiology. 2025 Mar 1;22(3):249-59.
78. Yu PL, Chiu KW, Lu J, Lui GC, Zhou J, Cheng HM, Mao X, Wu J, Shen XP, Kwok KM, Kan WK.
Application of a deep learning algorithm for the diagnosis of HCC. JHEP Reports. 2025 Jan
1;7(1):101219.
79. Sattari MA, Zonouri SA, Salimi A, Izadi S, Rezaei AR, Ghezelbash Z, Hayati M, Seifi M, Ekhteraei
M. Liver margin segmentation in abdominal CT images using U-Net and Detectron2: annotated dataset
for deep learning models. Scientific Reports. 2025 Mar 13;15(1):8721.
80. Zheng T, Zhu Y, Jiang H, Yang C, Ye Y, Bashir MR, Li C, Long L, Luo S, Song B, Chen Y. MRI‐Based
Topology Deep Learning Model for Noninvasive Prediction of Microvascular Invasion and Assisting
Prognostic Stratification in HCC. Liver International. 2025 Mar;45(3):e16205.
81. Zhang H, Liu J, Su D, Bai Z, Wu Y, Ma Y, Miao Q, Wang M, Yang X. Diagnostic of fatty liver using
radiomics and deep learning models on non-contrast abdominal CT. PloS one. 2025 Feb
13;20(2):e0310938.
82. Yu Y, Cao L, Shen B, Du M, Gu W, Gu C, Fan Y, Shi C, Wu Q, Zhang T, Zhu M. Deep Learning
Radiopathomics Models Based on Contrast-enhanced MRI and Pathologic Imaging for Predicting
Vessels Encapsulating Tumor Clusters and Prognosis in Hepatocellular Carcinoma. Radiology:
Imaging Cancer. 2025 Mar 14;7(2):e240213.
83. Gupta P, Hsu YC, Liang LL, Chu YC, Chu CS, Wu JL, Chen JA, Tseng WH, Yang YC, Lee TY, Hung
CL. Automatic localization and deep convolutional generative adversarial network‐based classification
of focal liver lesions in computed tomography images: A preliminary study. Journal of
Gastroenterology and Hepatology. 2025 Jan;40(1):166-76.
84. Lei Y, Feng B, Wan M, Xu K, Cui J, Ma C, Sun J, Yao C, Gan S, Shi J, Cui E. Predicting microvascular
invasion in hepatocellular carcinoma with a CT-and MRI-based multimodal deep learning model.
Abdominal Radiology. 2024 May;49(5):1397-410.
85. Duan T, Zhang Z, Chen Y, Bashir MR, Lerner E, Qu Y, Chen J, Zhang X, Song B, Jiang H. Deep
learning-based compressed SENSE improved diffusion-weighted image quality and liver cancer
detection: A prospective study. Magnetic Resonance Imaging. 2024 Sep 1;111:74-83.
86. Wei Y, Yang M, Zhang M, Gao F, Zhang N, Hu F, Zhang X, Zhang S, Huang Z, Xu L, Zhang F. Focal
liver lesion diagnosis with deep learning and multistage CT imaging. Nature communications. 2024
Aug 15;15(1):7040.
87. Wang F, Chen Q, Chen Y, Zhu Y, Zhang Y, Cao D, Zhou W, Liang X, Yang Y, Lin L, Hu H. A novel
multimodal deep learning model for preoperative prediction of microvascular invasion and outcome in
hepatocellular carcinoma. Eur J Surg Oncol. 2023;49(1):156–64. doi:10.1016/j.ejso.2022.08.036.
88. Kucukkaya AS, Zeevi T, Chai NX, Raju R, Haider SP, Elbanan M, et al. Predicting tumor recurrence
on baseline MR imaging in patients with early-stage hepatocellular carcinoma using deep machine
learning. Sci Rep. 2023;13(1):7579. doi:10.1038/s41598-023-34439-7.
Page 1378
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue II, February 2026
89. Stollmayer R, Budai BK, Rónaszéki A, Zsombor Z, Kalina I, Hartmann E, et al. Focal Liver Lesion
MRI Feature Identification Using EfficientNet and MONAI: A Feasibility Study. Cells.
2022;11(9):1558. doi:10.3390/cells11091558.
90. Yang Y, Zhou Y, Zhou C, Ma X. Deep learning radiomics based on contrast enhanced computed
tomography predicts microvascular invasion and survival outcome in early stage hepatocellular
carcinoma. Eur J Surg Oncol. 2022;48(5):1068-1077. doi: 10.1016/j.ejso.2021.11.120.
91. Anderson BM, Rigaud B, Lin YM, Jones AK, Kang HC, Odisio BC, Brock KK. Automated
segmentation of colorectal liver metastasis and liver ablation on contrast-enhanced CT images. Front
Oncol. 2022;12:886517. doi:10.3389/fonc.2022.886517
92. Zheng R, Wang L, Wang C, Yu X, Chen W, Li Y, Li W, Yan F, Wang H, Li R. Feasibility of automatic
detection of small hepatocellular carcinoma (≤2 cm) in cirrhotic liver based on pattern matching and
deep learning. Phys Med Biol. 2021;66(8):10.1088/1361-6560/abf2f8. doi:10.1088/1361-6560/abf2f8
93. Jiang YQ, Cao SE, Cao S, Chen JN, Wang GY, Shi WQ, Deng YN, Cheng N, Ma K, Zeng KN, Yan XJ,
Yang HZ, Huan WJ, Tang WM, Zheng Y, Shao CK, Wang J, Yang Y, Chen GH. Preoperative
identification of microvascular invasion in hepatocellular carcinoma by XGBoost and deep learning. J
Cancer Res Clin Oncol. 2021;147(3):821–833. doi:10.1007/s00432-020-03366-9
94. Liu Y, Ning Z, Örmeci N, An W, Yu Q, Han K, Huang Y, Liu D, Liu F, Li Z, Ding H, Luo H, Zuo C,
Liu C, Wang J, Zhang C, Ji J, Wang W, Wang Z, Wang W, et al. Deep Convolutional Neural Network-
Aided Detection of Portal Hypertension in Patients With Cirrhosis. Clin Gastroenterol Hepatol.
2020;18(13):2998–3007.e5. doi:10.1016/j.cgh.2020.03.034
95. Kumar S, Rani P. An AI-based liver disease prediction model based on Pearson correlation feature
selection method. Biomed Pharmacol J. 2024;17(4):2187–2202. doi:10.13005/bpj/3016.
96. Dosovitskiy A, et al. An image is worth 16x16 words: Transformers for image recognition at scale.
arXiv preprint arXiv:2010.11929. 2020.
97. Chakraverti S, Sharma S, Shukla V. Extensive analysis of deep learning algorithms for automatic
detection of liver anomalies using MRI and CT scanned images. 2024:1402–1409.
doi:10.1109/ictacs62700.2024.10841192.