![CNN Architecture for tongue image classification[20]](https://www.researchgate.net/profile/Ghazwan-Hani/publication/389857425/figure/fig3/AS:11431281315950903@1742014907961/CNN-Architecture-for-tongue-image-classification20_Q320.jpg)
Content uploaded by Ghazwan Hani
Author content
All content in this area was uploaded by Ghazwan Hani on Mar 15, 2025
Content may be subject to copyright.
Iraqi Journal of Intelligent Computing and Informatics (IJICI)
Vol. 4, 1, June 2025, pp. 33~56
ISSN: 2791-2868, DOI: 10.52940/ijici.v4i1.90 r 33
Journal homepage: https://ijici.edu.iq
A Comprehensive Review of Advances in Tongue Image
Classification Techniques for Diabetes Identification
Ghazwan H. Hussien1, Zainab N. Nemer 2
1,2 Department of Computer Science, Basra University, Basra, Iraq
Article Info
ABSTRACT
Article history:
Received January 31, 2025
Revised March 5, 2025
Accepted March 12, 2025
By grouping diabetics using tongue scans, therefore enabling
non-invasive, economically priced, and efficient approaches to
disease detection. Mostly focused on patient healthcare using
medical diagnostics and early detection, research has evolved. This
study issue has become more important since it supports early
diabetes detection, helps clinicians and patients, and targets proactive
treatments meant to reduce the condition. It helps doctors decide
which important diabetes treatments to use because this metabolic
disorder can damage many organs if it is not treated properly. Deep
learning algorithms have made it possible to diagnose many diseases
early, including diabetes, by processing and analyzing images of the
tongue to classify diabetic patients. This makes it possible to
combine feature extraction and pattern recognition. There are more
people with diabetes around the world, and the number of new cases
is also going up. People want accurate and reliable diagnostic tools,
so they've made algorithms that look at pictures of tongues and get
basic information from them. Using tongue photos, this paper
presents a comprehensive review of current advancements in the
classification and diagnosis of diabetes by focusing on developments
in deep network designs, feature extraction techniques, assessment
methods, and deep learning methodology. The approach uses tongue
images to routinely analyze frequently used datasets and indicators
for diabetes classification. It also covers the challenges faced and
perhaps the routes of research to provide innovative ideas in this
field.
Keywords:
Diabetes
Deep learning
Neural network
Tongue images
Corresponding Author:
Ghazwan H. Hussien
Zainab N. Nemer
Department of Computer Science, Basra University, Basra, Iraq
Email: ghazwan.hani@uobasrah.edu.iq
Email: zainab.nemer@uobasrah.edu.iq
1. INTRODUCTION
One of the most deadly diseases worldwide, diabetes affects a growing daily count of people. Diabetes
can be of two main types: Type I, sometimes known as insulin-dependent diabetes, results from the pancreas
not producing the vital hormone insulin required for survival. Despite its prevalence in children and
teenagers, this type primarily manifests in later life. With about 90% of all diabetes cases worldwide, the
second type—called non-insulin-dependent—results from the body's inadequate response to insulin
generated by the pancreas. Usually involving patients visiting a diagnostic center or doctor, the identification
process takes extra time and costs resources to produce their diagnosis. Therefore, the accurate and quick
diagnosis and analysis of diabetes is a topic deserving of research.[1] [2]
Diabetes mellitus (DM) has become a major public health issue for adults due to the rapid rise in
population growth, aging, obesity, and sedentary lifestyles. According to an epidemiological analysis on the
worldwide prevalence of diabetes, 171 million people had the condition worldwide in 2000; estimates of the
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
34
number by 2030 point to 366 million. Higher blood glucose levels and prolonged diabetes run the risk of
macrovascular or microvascular effects. Also, most people with diabetes had oral problems like gum disease,
tooth loss, dry mouth, cavities, burning mouth syndrome, problems with taste and salivary glands, slow
wound healing, lichen planus, geographic tongue, and candidiasis. People with diabetes mellitus will easily
show buccal changes. [3].
With the main purposes being pulse assessment and tongue examination, diagnosis in Traditional
Chinese Medicine (TCM) depends on four diagnostic techniques: observation, auscultation and olfaction,
inquiry, and palpation. Studies of tongue pictures in Traditional Chinese Medicine (TCM) show a strong
correlation with diabetes.[4].
The digital tongue tool TDA1 captures tongue images by establishing a stable light source environment.
These images primarily concentrate on two components: the tongue body and the tongue layer.
Distinguishing between these two elements is crucial for tongue diagnosis, as it facilitates hypotheses
regarding color and texture analysis. The division and merging technique and the color threshold method
were used to tell the difference between the tongue body and the tongue layer. This made it possible to get
parameters for each part. [4].
With tongue diagnosis being a fundamental component of Traditional Chinese Medicine (TCM), which
has a rich historical basis and is accepted as a common alternative medicine in Western countries. Through
its many properties—color, texture, and shape—the tongue becomes a diagnostic instrument for disorders.
Figure 1 shows the several tongue parts corresponding to the purposes of several body organs: The tip is
connected with the lungs, heart, chest, and neck; the central portion relates to the liver, spleen, stomach, and
pancreas; and the posterior section shows the abdominal organs. The small intestine and the colon (large
intestine) make up the lower intestine, so the system must focus on a certain location of the tongue since this
affected area indicates a malfunction in a given part of the body. [5].
We extract characteristics from all tongue photos, primarily focusing on the middle region due to its
connection to the pancreas. When doctors look for illnesses, they usually check the patient's health by
looking at the shape, color, and size of the tongue, since changes in any of these factors can reliably point to a
problem inside the body. In this proposed work, we develop a computer-based automated method to analyze
alterations in the tongue, which will subsequently aid in identifying diabetes in patients.[5]
Deep learning algorithms and deep convolutional neural networks (CNNs) have come a long way in recent
years. This has led to better classification accuracy and efficiency in image analysis technologies that use
CNNs. Originally used widely in picture segmentation, image classification, and face recognition,
convolutional neural networks (CNNs) have since become a key focus of study in the field of objective
tongue diagnosis. CNN architectures can independently extract features, therefore removing the necessity for
human feature selection—which is crucial for including intelligent tongue diagnosis systems into TCM
clinical practice.[6]
This study looks into how machine learning and deep learning can be used to classify and find people with
diabetes using pictures of their tongues. It also looks at how the system was made and highlights
improvements in network architectures, feature extraction methods, and assessment metrics. The study talks
about the problems this field is facing and suggests ways to do more research to make these systems more
accurate and useful in real-life healthcare situations.
Figure 1. Structure of tongue.
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
35
Previous research revealed strengths in the method of classifying diabetes based on tongue
images.The newest machine learning and deep learning technologies were combined with a strong
background in traditional Chinese medicine (TCM), specifically the use of images of the tongue to diagnose
diseases. This is a non-invasive and relatively easy way to collect data that doesn't require complicated
medical equipment or painful procedures for the patient. This combination opens up new possibilities for
diagnosing diseases in non-invasive ways and allows for automation and improvements to the diagnostic
process. The studies focused on linking a specific area of the tongue (the middle area) to diabetes based on
the principles of traditional Chinese medicine (its connection to the pancreas), increasing the accuracy of
diagnosis. This identification reduces the "noise" in the data and allows focusing on the features most
relevant to the disease.Deep Convolutional Neural Networks (CNNs) have proven to be very effective in
image analysis. They are able to automatically extract distinctive features from images (such as color,
texture, and shape) without the need for significant human intervention. This means reducing reliance on
human expertise in identifying important features (avoiding manual feature selection), making the process
more objective and repeatable. This leads to a faster and less expensive way to diagnose diabetes compared
to traditional methods (which rely on a doctor’s visit and a laboratory).
Given all the aforementioned, this study tackles the vital demand for ongoing diabetes diagnosis
improvement. This paper primarily focuses on enhancing the accuracy, dependability, and cost-effectiveness
of a non-invasive diagnostic system for early and accurate diabetes diagnosis, using tongue pictures and deep
learning approaches. While tongue image analysis is a useful substitute, current methods still need more
work and improvement to meet present difficulties. Rather than suggesting a whole new solution, this review
paper seeks to synthesize and assess current solutions, pinpoint best practices, and point up areas for future
research.
2. RELATED WORKS
This section analyzes prior studies on the classification and early detection of diabetes through tongue
imagery. This paragraph is to succinctly summarize many surveys on this topic to contextualize the current
research within the existing literature.
2020 saw Thirunavukkarasu et al. look at the viability of using tongue thermography as a non-invasive
early type II diabetes diagnosis tool. The study, which involved 140 subjects—70 healthy individuals and 70
diabetics—investigated the heat distribution in the tongue region using image processing techniques and
machine learning approaches. The results showed that the tongue's surface temperature in diabetics was
significantly higher than that of normal participants, showing a strong statistical link between glycemic
haemoglobin (HbA1c) levels and thermal distribution in the tongue region (r² = 0.5688). Attaching the
highest classification accuracy at 94.28%, the Convolutional Neural Network (CNN) approach The study
found that tongue thermal imaging could be a useful, non-invasive approach for early type II diabetes
diagnosis. [7].
Sagayaraj et al. (2021) investigated the use of tongue photos for the detection of diabetes and
diabetic retinopathy using image processing and machine learning methods. The study sought diabetic
patients using tongue features including color, texture, and shape. The Bi-Elliptical Deformable Contour
(BEDT) technique helped to separate images and retrieve pertinent elements. A Support Vector Machine
(SVM) splits images into two groups: healthy persons and those with diabetes. The findings revealed a
decent classification accuracy—that is, an accuracy rate of 88.28% in diabetic identification. [8].
Li et al. (2022) investigated the distribution of tongue characteristics in diabetics by using
unsupervised learning methods. The study's goal was to find out the rules that govern how the different parts
of a diabetic's tongue are distributed. This was done so that traditional Chinese medicine (TCM) principles
could be used to create a diagnostic base for personalized diabetes treatment. The TFDA-1 tongue diagnostic
device was used to get images of the tongues of 598 patients, and the Tongue Diagnostic Analysis System
(TDAS) was used to measure chromatic, tactile, and layer ratios. K-means and self-organizing maps (SOM)
algorithms were employed to examine the distribution of tongue features. The findings indicated that the K-
means algorithm categorized patients into three clusters, but the SOM method classified patients into four
clusters. We identified statistically significant changes in chromatic and tactile attributes among groups,
suggesting that these features could serve as precise indicators of patients' conditions. The research
determined that employing unsupervised learning methods may effectively identify small alterations in
tongue characteristics among diabetics, facilitating a more precise diagnosis and treatment. [9].
In 2022, Balasubramaniyan et al. developed a diagnostic model for diabetes employing panoramic
tongue imaging techniques and deep learning. The researchers employed a convolutional neural network
(CNN) featuring a ResNet-50 architecture to examine tongue attributes, encompassing color, texture, shape,
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
36
and dental imprints. The data is categorized using the Deep Radial Basis Function Neural Network (RBFNN)
method through automated learning. The model exhibited outstanding performance, with a diagnostic
accuracy of 98.4%, exceeding that of rival models such as ResNet-34 and AlexNet. The study concluded that
this non-invasive method could be an excellent tool for the early detection of diabetes [10].
The latest models employing machine learning and deep learning for diabetes detection have been
analyzed (Wee et al., 2024). The study investigated the influence of redundant sampling techniques and
feature selection on model effectiveness, highlighting the use of non-invasive and anthropometric
measurements to improve diagnosis accuracy. The findings demonstrated that deep learning models,
encompassing convolutional neural networks (CNN) and deep neural networks (DNN), outperformed
traditional machine learning models like random forests (RF) and logistic regression (LR). The study
revealed issues related to data availability and quality, emphasizing the need for data-driven methodologies
to analyze medical factors for improved diagnostic reliability. [11].
This study, conducted by Usharani Thirunavukkarasu and colleagues and published in 2024, aimed
to develop a non-invasive method for diagnosing type II diabetes by combining thermal and visual pictures of
the tongue using discrete wavelet transform (DWT). Statistical features were extracted from embedded
images using the GLCM algorithm, and the data was classified employing machine learning classifiers
(SVM, LDA, K-NN) and deep learning models (VGG16, ResNet50). The results demonstrated that the mean-
max waveform conversion base had enhanced performance, exceeding VGG16 with an accuracy of 94.37%.
The results suggest that the combination of thermal and visual images may be an effective tool for the early
detection of diabetes. [12].
Burcu Tiryaki and colleagues conducted research on tongue lesions using medical imaging using
deep convolutional neural networks (DCNNs) and published it in 2024. Six thousand tongue images were
arranged into five groups: natural tongue, covered tongue, geographic tongue, cracked tongue, and median
rhombic glossitis. With a majority voting system to improve outcomes, the performance was evaluated using
VGG19, ResNet50, ResNet101, and GoogLeNet networks. The results showed that whilst VGG19 achieved
an accuracy of 83.93% in multi-class classification, ResNet101 achieved the best accuracy of 93.53% in
binary classification (normal vs. pest. Following the majority vote strategy improved accuracy to 95.15% and
88.76%, respectively. The results imply that this approach is relevant in clinical environments for tongue
lesion diagnosis. [13].
The associated works clearly show some restrictions and disadvantages. Many researchers suffer
from a limited comparison between several approaches, usually concentrating on a small number of
algorithms or structures. Furthermore, a regular difficulty is the interpretability of results, particularly for
advanced methods such as deep learning. Certain techniques depend on hand-crafted feature extraction,
which calls for knowledge, might not be best for catching intricate patterns, and can be arbitrary. Deep
learning techniques—including CNNs—extract features automatically, therefore providing possible benefits
in terms of power and flexibility. Different research, meanwhile, employs various kinds of data. While some
techniques involve infrared imaging, others use plain, visible tongue images. Thirunavukkarasu et al. point
out that thermal imaging can be sensitive to outside variables such as room temperature and recent meal or
beverage intake. A few researchers, like Li et al., examine tongue feature distributions using unsupervised
learning (K-means, SOM), which does not directly categorize patients but offers an understanding of feature
patterns within a known diabetes group. Understanding the outcomes of unsupervised learning can prove
difficult. Other constraints include the reliance on quite small datasets, the complexity brought about by
combining several techniques (e.g., merging thermal and visual pictures and employing different classifiers),
and the usage of panoramic tongue images—which may not always be available.
3.DISCUSSION
Analyzing a range of studies on tongue image classification for diabetes detection, this paper
emphasizes the growing interest in combining Traditional Chinese Medicine (TCM) concepts with modern
machine learning and deep learning approaches. From this research, a number of significant trends and
observations emerge. First of all, deep learning models—especially convolutional neural networks (CNNs)—
clearly are being embraced given their ability to automatically extract relevant features from complex visual
data. Effective applications of architectures such as ResNet, VGG, DenseNet, and MobileNet illustrate the
capacity of deep learning to outperform conventional techniques dependent on hand-crafted feature
extraction. Second, while much research focuses simply on visual tongue photographs, some have looked at
the benefits of using thermal imaging and suggest that multimodal techniques can improve diagnostic
accuracy. Thirdly, as ensemble learning techniques (bagging, boosting, stacking, and voting) are being used
more and more, model performance and robustness are improving.
Still, various constraints and difficulties exist. Small sample sizes and a lack of diversity in the
datasets utilized in many of the examined research cause questions regarding the generalizability of the
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
37
results. Furthermore, impeding repeatability and comparison among investigations is the reliance on
particular databases, usually not publicly accessible. Moreover, even if CNNs show potential, their "black
box" character makes it challenging to understand why a given categorization judgment is taken, which is
essential for developing confidence and acceptability among medical practitioners. Methods based on, say,
decision trees have less of the interpretability problem, but they usually compromise accuracy.
Standardization of image acquisition presents still another difficulty. Classification model performance can
be much influenced by changes in camera angles, picture resolution, and lighting conditions. Though
referenced in many publications, the TDA1 instrument is a step towards standardization; wider acceptance
and validation are still needed. Ultimately, a crucial factor for practical application is that few research
studies specifically address the ethical and privacy issues related to gathering and using medical picture data.
Larger, more varied, publicly accessible datasets of tongue pictures with standardized image capture
techniques should take center stage in future studies. Explainable artificial intelligence (XAI) methods have
more to be worked on to make deep learning models more interpretable and transparent. Another exciting
path for enhancing diagnostic accuracy and offering a more complete evaluation of patient health is the
combination of tongue image analysis with other clinical data (e.g., patient history, blood test findings).
Before these approaches may be generally embraced in clinical practice, thorough clinical validation studies
are finally indispensable to show their real-world efficacy and safety.
4. RECOGNITION ON CLASSIFICATION MODEL
Tongue scans of both diabetic and non-diabetic groups yielded the data, mostly related to color and
textural characteristics. Preprocessing included independent variables derived from the feature parameters.
The categorized variable was considered to be either diabetic or absent (dependent variable).Twenty percent
of the specimens were set aside for testing, while eighty percent were used for training. Figure 2 depicts the
model.[4] complete tongue image classification system, the different stages of the data starting from training
data, through feature extraction (tongue color, tongue texture, etc.), feature processing (association
processing, sample equalization, feature normalization, feature reduction), classifier training and
optimization, and finally performance evaluation on independent test data. Each of these stages plays a
crucial role in achieving high accuracy and reliability.
The disparity in sample sizes affects the classification model; therefore, to equalize the samples between
the diabetic and non-diabetic groups, the Synthetic Minority Over-sampling Technique (SMOTE) was
employed. The augmentation of variables would render the classification model more intricate; thus,
Principal Component Analysis (PCA), a traditional method for dimensionality reduction, was employed to
process the raw features obtained from tongue images, ensuring the preservation of information integrity
during the dimensionality reduction procedure [4].
5. PIPELINE OF TONGUE IMAGE CLASSIFICATION
Timely detection of diabetes is crucial; however, depending just on traditional surgical methods
based on blood tests is insufficient and requires the acceptance of modern non-surgical alternatives. With
benefits in comfort and patient convenience, tongue image classification is a new non-invasive approach for
early diabetes detection. The approaches and techniques applied in tongue image classification for several
machine learning and deep learning systems are investigated in this part.
5.1. Deep Learning and Machine Learning
Figure 2. Map of classification model
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
38
While deep learning is a type of machine learning that uses neural networks to emulate the human
brain's learning process, in artificial intelligence (AI), machine learning refers to the ability for autonomous
adaptation with minimal human intervention. These two ideas differ really significantly. Deep learning can
adapt to new circumstances and correct its own mistakes even if training requires more data. On the other
hand, machine learning lets one learn on smaller datasets; yet, it depends on more human engagement to
learn and fix its mistakes. [14].
Based on tongue images, several well-known approaches have been suggested for the diagnosis and
classification of diabetes or non-diabetic conditions, including machine learning techniques such as Gradient
Boosting (GB), Support Vector Machine (SVM), AdaBoost (AB), Random Forest (RF), Decision Tree (DT),
Naive Bayes, and deep learning techniques.[15]. Tongue image research Artificial intelligence (AI), machine
learning (ML), and deep learning (DL) drive non-invasive diabetes screening. An area of machine learning,
deep learning independently extracts and learns intricate information from medical images using several
hidden layers of deep neural networks (DNNs). For the analysis of tongue pictures, where minute variations
in form, color, and texture could indicate diabetes progression, these characteristics are crucial. [12].
On the other hand, developments in deep learning—particularly in relation to 2010—provided
convolutional neural networks (CNNs), which greatly exceeded traditional approaches. These models
independently master hierarchical properties from data. Fused thermal and visual tongue images were used to
distinguish diabetic from non-diabetic patients utilizing networks including VGG16 and ResNet50. These
models achieved classification accuracy of 94.37% (VGG16) and 78% (ResNet50) by combining low-level
properties (e.g., edges, color) with high-level abstractions (e.g., patterns, textures), using fused images. This
stands in stark contrast to the accuracy of individual modalities, such as visual images (85%) and thermal
imaging (90.62%), therefore stressing the effectiveness of image fusion combined with deep learning. [12].
The application of deep learning in tongue image processing has revolutionized the diagnosis of
diabetes mellitus (DM), offering a non-invasive and efficient alternative to traditional methods. Deep
learning models, particularly Convolutional Neural Networks (CNNs), are particularly adept at autonomously
extracting high-level features from photographs, such as texture, color, and geometry. This approach
provides a more comprehensive perspective than conventional handmade methods.This method emphasizes
the potential of deep learning to transform healthcare diagnostics by enabling the creation of instruments that
are scalable, portable, and precise, thereby facilitating the early detection of diabetes. By integrating transfer
learning, CNNs, and data augmentation, this approach creates a strong foundation for future advancements in
tongue-based diagnostic systems. [15].Figure 3 shows the diabetes detection system using tongue images.
The system starts by taking the patient's tongue image, then processing it and improving its quality using an
adaptive median filter, then the image is segmented using a deep neural network (ResUNet++) to determine
the part that contains only the tongue image, then the image is amplified by making slight changes to the
image to create new images and increase the number of images, which helps in reducing the excess
oversampling, then the features are extracted from the image using a convolutional neural network, and
diabetes is detected using another deep network (deep residual network) used for classification.[14].
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
39
Figure 3. DM detection using tongue image
5.2. Convolutional Neural Networks
Image classification, recognition, and object detection challenges have shown remarkable
performance of Convolutional Neural Networks (CNNs). Designed to allow for shifts, scales, and distortions,
a standard CNN consists of input, convolutional, pooling, fully connected, and output layers. They are adept
at spotting intra-class differences in texture, posture, and lighting as well as in changes in stance.
Emphasizing factors including color, texture, and fissures, CNN-based models have been trained on large
datasets to classify diabetes through the study of tongue images. These models fit for medical diagnosis since
they show resistance to changes in tongue form and illumination. Still, their effectiveness depends much on
data quality; therefore, exact results depend on high-resolution photos. CNNs are quite good at spotting
complex patterns, but generalization across different datasets and model interpretability still suffer. Later
studies should give top priority to improving data quality and developing explainable artificial intelligence
techniques to increase their dependability in medical use.[15] [16].
Common architectures employed in tongue image classification for diabetes detection include
DenseNet, GoogleNet, and MobileNet. These designs, despite utilizing relatively little datasets, effectively
reduced mistakes and decreased training durations by employing efficient parameter sharing, streamlined
design, and integrated graphics processing units (GPUs).
Numerous CNN-based architectures, including VGG16, VGG19, and ResNet50, are extensively
employed in computer vision tasks, such as tongue image categorization for diabetes diagnosis. These models
provide neural representations that efficiently extract both low- and high-level features from tongue pictures,
enhancing diagnostic precision. Notwithstanding the attainment of considerable accuracy (e.g., VGG16
achieving 94.37% post-image fusion), their complexity and prolonged training duration persist as key
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
40
obstacles. As tongue image analysis has advanced, the utilization of deeper neural networks such as
ResNet50 has become essential. Nevertheless, augmenting the number of layers in neural networks presents
challenges in training and may sometimes result in diminished accuracy due to phenomena such as
overfitting or vanishing gradients. Utilizing approaches such as transfer learning and image fusion (e.g.,
integrating thermal and visual tongue images) alleviates these issues, enhancing performance and resilience
in diabetes detection. Figure 4 shows the architecture of the modified VGG16 model. It takes a standard-
sized tongue image (224 × 224 pixels) on which the model was trained and with a depth of 3 (representing
the color channels: red, green, and blue—RGB), the five convolutional blocks extract features from the
image with the help of filters of different sizes, each of which produces its own feature map, and the ReLU
activation function helps the network learn non-linear relationships between features, then converts the
outputs of the convolutional blocks (which are three-dimensional matrices) into one long vector (one-
dimensional). This is necessary to prepare the data for input to the fully connected layers; then the fully
connected layers learn the relationships between the features extracted in the previous stages, applying the
SoftMax activation function in the last layer to convert the outputs (which are numbers) into probabilities.
There is a probability for each class (in this case, "Normal" and "Diabetes"); the sum of the probabilities is
equal to 1, and the class with the highest probability is the class that the network predicts as the final
prediction.[12].
The residual network (ResNet), developed to improve performance and accuracy through the
incorporation of deeper layers, has been modified for tongue image classification to differentiate diabetic
patients from healthy ones. The supplementary layers of ResNet enhance feature extraction, although they
necessitate empirical assessment to confirm they do not detrimentally impact model performance.
Lightweight deep neural networks, such as MobileNet, are particularly effective for tongue image processing
due to their efficient computational capabilities and strong performance with optimized hyperparameters.
[17]. Prominent CNN-based architectures like Inception and its variants demonstrate innovation through
modular designs that enhance feature extraction in intricate images. These modules proficiently analyze
texture, color, and geometric attributes vital for diagnosing diabetes. Xception, an advanced version of
Inception, employs depth-wise separable convolutions, improving computational efficiency and classification
accuracy for combined thermal and visual tongue data.[18]. CNN-based models' adaptability emphasizes
their ability to improve non-invasive diabetes detection using tongue images.[19]. Figure 5 shows a
conceptual diagram of a CNN-based system for classifying diabetic tongues by passing the tongue image
through a series of convolutional layers plus an activation function (ReLU) to extract features from the image
(from simple to complex), then a pooling layer to reduce the data size and make the network more resistant to
minor changes, then flatten to convert matrices to a vector, then fully connected to learn the relationships
between the features, then softmax to output probabilities for each class, and select the class with the highest
probability.
Figure 4. Pre-trained VGG-16 Net for tongue image classification.
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
41
Figure 5. CNN Architecture for tongue image classification[20]
5.3. Ensemble learning
Ensemble learning is a crucial method in the field of machine learning. In recent years, ensemble
learning has garnered significant attention in artificial intelligence, pattern recognition, machine learning,
neural networks, and data mining. Ensemble learning has proven to be effective and practical across a broad
range of problem domains and significant real-world applications. Ensemble learning generates many
classifiers or a set of base learners and amalgamates their outputs to diminish overall variance. The
amalgamation of many classifiers or base learners markedly enhances accuracy compared to a solitary
classifier or base learner [21].
By aggregating predictions from many base models, ensemble learning techniques have achieved
exceptional performance in many machine learning applications. From their inception to modern state-of-the-
art algorithms, this paper offers a concise overview of ensemble learning covering the three main ensemble
methods: bagging, boosting, and stacking. Previous studies have focused mostly on common ensemble
techniques, including both machine learning and deep learning approaches. Machine learning, a subset of
artificial intelligence (AI), has advanced thanks to the many discoveries and developments in many spheres
of research, including aspects of human existence, in recent years. Machine learning can also be difficult at
the same time, especially with unbalanced and high-dimensional datasets. As such, researchers regularly
apply fresh and improved learning strategies, including ensemble learning. [22].
5.3.1. Ensemble learning techniques Basic
Ensemble learning is a powerful paradigm in machine learning where multiple models, often called
“base learners” or “weak learners,” are strategically combined to solve a given problem. The basic idea is
that by combining the predictions of multiple models, the ensemble can achieve better predictive
performance than any of its individual component models. This works best when the base learners are diverse
(i.e., they make different types of errors). Common ensemble learning techniques include soft voting and
majority voting.
5.3.1.1. Averaging (soft voting)
Using columns for the models and rows for the predictions for every data sample, soft classification
arranges the computed probabilities by each ensemble model on the validation data into a matrix. For every
sample, the arithmetic mean of the model predictions is calculated; the class with the highest cumulative
probability is then chosen. The potential of overfitting is a main drawback. Soft voting can produce less-
than-ideal performance if the underlying models are overfitted. Exhibit soft voting; Illustration 6. [23].
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
42
Figure 6. Show soft voting
# Pseudo code (Soft Voting)
fold_predictions = []
val_predictions = predict_probabilities(model, X_val)
append(fold_predictions, val_predictions)
fold_predictions = convert_to_3D_array(fold_predictions)
average_probabilities = calculate_average(fold_predictions, across_models)
final_predictions = get_class_with_highest_average_probability(average_probabilities)
5.3.1.2. Max voting (Majority voting)
Majority voting ranks every class across all models and then determines the class with maximum
prediction frequency. This yields a more realistic assessment of the model's performance. Majority voting
works on a matrix whereby rows indicate models and columns represent classes or samples. Figure 7 shows
overall voting.[24].
Figure 7. Show majority voting
# Pseudo code (Majority Voting)
FUNCTION majority_vote(predictions):
num_models = get_number_of_models(predictions)
num_samples = get_number_of_samples(predictions)
final_predictions = []
FOR EACH sample_index FROM 0 TO num_samples - 1:
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
43
sample_predictions = get_predictions_for_sample(predictions, sample_index)
most_frequent_class = find_most_frequent(sample_predictions)
append(final_predictions, most_frequent_class)
final_predictions = convert_to_array(final_predictions)
RETURN final_predictions
5.3.2. Ensemble learning techniques advanced
Among advanced ensemble learning methods are bagging, boosting, and stacking. These techniques
extend on the fundamental ideas of averaging and voting but provide more complex ways to generate and
combine base learners.
5.3.2.1. Stacking
Stacking is the method of integrating various estimators to mitigate their biases. The forecasts from
each estimator are aggregated and utilized as input for a final estimator, commonly referred to as a meta-
model, which generates the ultimate prediction. The ultimate estimator is trained by cross-validation.
Stacking is applicable to both regression and classification tasks. Figure 8. Illustration of the stacking
technique. By training several different machine learning models (C1, C2, C3) on the training data, we use
these models to make predictions (P1, P2, P3), train another model (the super-classifier) on the relationship
between these predictions and the correct classification, and we use the super-classifier to make the final
prediction.[25].
Figure 8. Show stacking technique [25].
Stacking begins with the division of data into a training set and a validation set, followed by the
partitioning of the training set into K folds (for instance, 5 folds). The base model is trained on 4 folds while
predictions are generated for the fifth fold, and this process is reiterated until predictions are obtained for
each fold. Train the base model on the complete training set, utilize the model to generate predictions on the
test set, iterate through the process from training the model on the remaining folds to predicting on the test
set, and employ the test set predictions as features for a new model—the meta-model—before ultimately
producing final predictions for the test set using the meta-model. In regression tasks, the inputs to the meta-
model are numerical, but in classification tasks, they consist of probabilities or class labels [26].
5.3.2.2. Bagging
Obtaining random data samples, using learning algorithms, and using the mean help one to find
bagging probabilities. Called aggregating bootstraps, the bagging process combines results from several
samples to generate an overall output. Bagging is creating many subsets of the original dataset with
replacement, building a base model for every subset, running all models concurrently, and aggregating the
predictions from every model to generate the ultimate predictions.[27].
5.3.2.3. Boosting
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
44
Boosting functions by integrating a sequence of weak learners, assigning greater weight to faulty
predictions in succeeding iterations and lesser weight to correct predictions. This compels the algorithm to
concentrate on observations that are challenging to anticipate. The ultimate prediction derives from the
majority or aggregate of the votes. Boosting can be employed to address regression and classification
challenges. The "Boosting Classifier" is employed to ascertain the quantity of weak learners in the ensemble,
denoted as `estimators_n. while the influence of each weak learner on the final cluster is regulated by `rate
learning'. [28] [29].
5.3.2.4. Comparison between Bagging, Boosting and Stacking
Among the most important methods used in machine learning to increase model accuracy and
reduce errors are ensemble learning methods. These approaches, which all rely on the idea of combining
several models to improve performance relative to using a single model, include bagging, boosting, and
stacking. Table 1 compares these three approaches with regard to their goals, training strategies, diversity,
benefits, and drawbacks, thereby helping to clarify their features and support the choice of the most suitable
way for the given work.
Table 1: Comparison between stacking, boosting and bagging [30].
Standard
Bagging
Boosting
Stacking
Objective
Reduce variance
Reduce bias
Improve overall
performance
Training
method
Parallel independent models
Sequential error-based
models
Various models + meta
model
Diversity
Depends on sample diversity
Depends on error correction
Depends on basic model
diversity
Advantages
Reduces variance, easy to
apply
Reduces bias, high
performance
High flexibility,
excellent performance
Disadvantages
Less effective at reducing bias
Prone to over fitting
More complex
5.4. Generative Adversarial Networks
Generative Adversarial Networks (GANs) may independently deduce and integrate fundamental
patterns from incoming data without requiring highly annotated training datasets. A Generative Adversarial
Network (GAN) comprises two neural networks: a generator and a discriminator. [31]. By means of
stochastic data—more especially, random values obtained from a preset distribution—the generator generates
new features. Acting as a binary classifier, the discriminator assesses the genuineness of the produced
features as either real or fake.[32].
The "adversarial" feature results from the fact that in the competitive training paradigm of GANs,
adversarial loss functions are maximized in a minimax game between the discriminator and generator. Two
among the several medical uses of Generative Adversarial Networks (GANs) are creating lifelike medical
images and improving diagnosis accuracy.[31]. Therefore, GANs can be used to improve the classification of
diabetic patients by adding extra realistic photographs for training, increasing the model's ability to identify
disease-related trends in tongue images.[33].
5.5. Machine learning
Machine learning classification algorithms are employed to classify a wide range of data. They
facilitate the categorization of commodities into similar or distinct categories based on their characteristics.
These algorithms are indispensable in the field of medical diagnostics, particularly in the context of image
classification tasks. This investigation enables the classification of tongue images and evaluates their
correlations with diabetes-related patterns. This exploratory study utilizes six unique conventional machine
learning classification methods to accurately identify diabetes patients through tongue images.[34].
5.5.1. K-Nearest Neighbors (KNN)
The quantity of nearest neighbors to be considered is denoted by "K" in the multidimensional
classification of nearest neighbors. KNN is widely employed in classification and regression problems due to
its simplicity, ease of implementation, and adaptability to nonlinear data. However, it encounters challenges
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
45
when dealing with large datasets, as the process of computing the distances between all data points can
impede performance. KNN's accuracy may be compromised by data and high-dimensional noise. [35].
KNN is used in medicine to classify people with diseases, including diabetes and heart disease, as
well as to project disease outcomes depending on symptomatology. It is quite effective in exercises on
classification and regression analysis. In one specific investigation, the KNN algorithm was applied to an
Alzheimer's MRI dataset for the disease diagnosis. The method showed reasonable performance with an
accuracy of 45.86%, therefore highlighting its potential in medical diagnostics despite natural
limitations.[36].
5.5.2. Decision Trees (DT)
A Decision Tree (DT) is a technique that employs a sequence of consecutive if-then principles to
classify input into classifications or judgments. It operates by identifying the most advantageous
characteristics to iteratively partition the data until categorical or predictive results are achieved. The decision
tree (DT) is a widely used tool in classification and regression problems, as it has the ability to manage both
numerical and categorical data and provides excellent interpretability. However, it is prone to overfitting if
the tree is extremely deep, and it may demonstrate sensitivity to minor data fluctuations, which could impact
the model's stability.[37].
Decision trees are flowchart-like structures that are used in medicine to make differential diagnoses
based on the patient's medical history and symptoms. They are particularly skilled in identifying risk factors
for conditions such as cardiovascular ailments and diabetes. The efficacy of DT models in clinical
environments has been demonstrated in previous studies. The DT model was trained on clinical data in one
investigation, which yielded an accuracy of 74.4%. The model's balanced accuracy was significantly
improved to 91.1% as a result of the integration of medical image feature analysis. The research concluded
that the integration of clinical data and medical picture elements by machine learning. [38].
5.5.3. Logistic Regression (LR)
The probability of data belonging to a specific category or a predetermined set of categories is
represented by the Logistic Regression (LR) algorithm. It is particularly well-suited for binary and linear
classification endeavors, as it functions primarily as a classification instrument rather than a regression
model, despite its designation. LR is one of the most widely used models in practical and commercial
applications due to its extraordinary efficiency and simplicity. However, its effectiveness is limited in
nonlinear tasks or when the interrelations among characteristics are complex, as it struggles to capture
complex patterns in these situations. [39].
In medicine, logistic regression is frequently employed to predict disease diagnoses, such as
diabetes, and is valued for its ability to provide probabilities alongside classifications. This makes it highly
advantageous for binary classification problems. A logistic regression model was developed and validated to
differentiate between peripheral lung cancer (PLC) and solitary pulmonary tuberculosis (SP-TB) using
clinical and imaging features. The model achieved a maximum AUC value of 0.878 in the internal validation
cohort, demonstrating its strong effectiveness in distinguishing between the two scenarios. A further study
utilized logistic regression to predict lymph node malignancy in pancreatic cancer based on ultrasound
imaging features, demonstrating it to be the most effective model. Additionally, logistic regression was
employed in a study to identify patients at heightened risk of kidney stones via CT-based radiography,
outperforming other machine learning models and underscoring its effectiveness in medical diagnostics. [40].
5.5.4. (Support Vector Machine - SVM)
Designed for binary classification and multi-class applications, the Support Vector Machine (SVM)
is a strong technique. It guarantees good generalization by finding the optimal hyperplane that divides data
into discrete categories with the biggest margin. Support vector machines (SVMs) help to construct linear
decision boundaries by nonlinearly translating inputs to high-dimensional feature spaces via kernel functions.
This approach reduces the risk of overfitting while yet making SVMs rather effective in managing both linear
and nonlinear data as well as high-dimensional datasets. Still, SVMs may show longer training times with
large datasets and need careful parameter fine-tuning, including kernel type and penalty coefficient (C), to
get the best performance.[33].
Support Vector Machines (SVMs) are widely utilized in the medical field for the classification of
medical images to detect various conditions, including diabetes, breast cancer, lung cancer, and osteoporosis.
Certain studies have employed SVMs to distinguish between the overall dimensions of the tumor and normal
liver tissue in patients with hepatocellular carcinoma. In a distinct experiment, the SVM model achieved an
outstanding 91.5% accuracy in classifying different types of brain tumors and distinguishing them from
normal cases. Additionally, support vector machines (SVMs) were utilized to predict BRAFV600E mutations
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
46
in patients with papillary thyroid cancer (PTC) using ultrasound, as part of a comparative study involving six
different machine learning models, with the SVM model exhibiting superior efficacy. These applications
highlight the versatility and effectiveness of SVMs in medical diagnosis and disease classification. [41].
5.5.5. Random Forest (RF)
The Random Forest (RF) algorithm is an ensemble technique that constructs several decision trees,
each trained on random subsets of data and attributes. Aggregating the outcomes of these trees facilitates
final classification or forecasting choices, thereby diminishing variability and improving accuracy. Random
forests are user-friendly, adept at managing extensive and intricate datasets, and exhibit a reduced
susceptibility to overfitting in comparison to solitary decision trees. Nevertheless, they may exhibit reduced
predictive speed when processing extensive information. [31].
Random forests comprise many decision trees that deliver robust performance and provide insights
on feature importance, making them highly valuable across various applications. In medical diagnostics, they
are widely employed for their accuracy and ability to manage large datasets with numerous variables. In a
study, six machine learning models were developed to distinguish between Epstein-Barr virus-related gastric
cancer (EBVaGC) and non-EBVaGC gastric cancer using CT radiography and clinical characteristics. The
Random Forest model outperformed the others, demonstrating reliable performance with high accuracy,
sensitivity, and specificity in the test dataset. A distinct study analyzing plantar dynamic pressure data to
detect osteoarthritis changes in the knee determined that Random Forest (RF) was the most successful model
among five options, which included K-Nearest Neighbors (KNN), Support Vector Machine (SVM),
AdaBoost, and XGBoost. Additionally, a study aimed at creating an AI-assisted model for diagnosing thyroid
disorders revealed that the Random Forest model exhibited greater accuracy compared to eight other machine
learning models, hence affirming its effectiveness in medical applications.[42].
5.5.6. Naïve Bayes (NB)
The Naive Bayes (NB) algorithm is a statistical classification technique that is based on Bayes'
theorem. It facilitates the rapid and efficient processing of high-dimensional data by operating under the
"naivety" assumption, which assumes that characteristics are mutually independent. This renders Naive
Bayes particularly pertinent for applications such as medical classification (e.g., disease identification) and
text classification (e.g., malware detection). The method is proficient in the management of missing data, is
easy to implement, and is efficient in the training and prediction processes. However, its precision may be
compromised if the independence assumption is violated or if the dataset is skewed.[43].
Built on Bayes' theorem, the Naive Bayes (NB) algorithm is a statistical categorization method.
Operating under the "naivety" assumption—that traits are mutually independent—it allows fast and effective
processing of high-dimensional data. Naive Bayes is thus particularly suitable for uses like text classification
(e.g., spam detection) and medical classification (e.g., disease diagnosis). The approach is easy to apply,
effective in training and prediction, and good in handling missing data. Still, its accuracy could drop if the
dataset is biased or the independence condition is violated. [44].
Six unique classic machine learning classification methods are delineated, alongside additional
methodologies capable of successfully identifying diabetic individuals through tongue image analysis. The
precision and efficacy of these algorithms differ from one to another.
5.6. Feature extraction
Diabetic categorization algorithms using tongue images depend critically on feature extraction.
Recognized as distinct traits that can point to the disease's presence are tongue pigment, scalp condition,
fissures, and other signs. Still, this method is complicated by issues like tongue discoloration from food dyes
or a white coating, which calls for changes to current systems to ensure exact and consistent feature
representation [45].
A deep learning method used for feature extraction, convolutional networks (CNNs) have been
applied to examine tongue photos and determine their correlation with either people diagnosed with diabetes
or those without the ailment. These networks detect complex patterns in tongue photos by using automatic
feature extraction through convolutional layers. Their training on datasets including photos of both diabetics
and non-diabetics helps them to better understand subtleties such as tongue color differences or the presence
of fissures. Realistic tongue images are created using generative adversarial networks (GANs), therefore
improving the variety of training material. This approach generates extra images that resemble real data,
therefore enhancing the model's ability to detect aberrant traits in tongue images, especially in situations with
little data. [46].
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
47
One of the difficulties with feature extraction is tongue contamination since food or beverage
pigments may affect its coloration, thereby complicating the extraction procedure. Furthermore, negatively
affecting model performance is the lack of data resulting from small available datasets. Furthermore, the
variability of features, which differs among individuals and cultural situations, needs models adept at
accepting this diversity. Proposed solutions to tackle the challenges include employing deep learning
methodologies such as ResNet, DenseNet, GoogleNet, and MobileNet to enhance model performance in
extracting intricate features from tongue images; augmenting data diversity through techniques like GANs to
generate realistic and varied tongue images; and utilizing ensemble learning strategies such as random forest,
stacking, boosting, bagging, average, and majority to elevate classification accuracy by consolidating the
outcomes of multiple models. Figure 9 illustrates the basic design of a CNN for feature extraction by
transforming the original image into a set of feature maps that represent increasingly complex and abstract
patterns.Layers of convolution pick filters to identify these trends. ReLU functions introduce nonlinearity.
Combining layers lowers dimensionality and strengthens the network's resistance to small variations.
Multiple consecutive layers of this process let the network learn a hierarchical representation of the image
(from simple features in the first layers to sophisticated features in the deeper layers).[47].
Figure 9. CNN Basic architecture for features extraction [20].
5.7. Data augmentation
One often used method to maximize the use of information gathering is data augmentation. Little
changes, such as scaling, rotating, and translating, help to improve the training set. These changes generate
fresh and diverse situations for training datasets, hence improving the efficacy and results of machine
learning models. Deep learning and other machine learning algorithms depend on large datasets for optimum
training; however, the restricted availability of datasets, particularly in emerging research areas, forces the
usage of data augmentation to increase the variety of present datasets by means of modifications.[48]
The effectiveness of data augmentation is derived from fundamental modifications, including
horizontal rotation, color space modification, and random cropping. These modifications encompass a
multitude of variables that pose challenges for applications that rely on image identification. Sophisticated
augmentation methods, in addition to fundamental approaches, include geometric modifications (such as
rotation, scaling, and translation), color space transformations (including adjustments to brightness, contrast,
and saturation), kernel filters (such as blurring or sharpening), and picture shuffling. Additionally, the
diversity of training data is being increasingly improved through the implementation of random erasure,
feature space augmentation, adversarial training, GAN-based augmentation, neural style transfer, and meta-
learning strategies. [49].
Combining several augmentation techniques increases the amount of the dataset and provides
several options that improve the durability and generalization of machine learning and deep learning models.
This is especially crucial in disciplines such as medical imaging, where datasets are usually restricted, and the
capacity to generalize across multiple circumstances is imperative for precise diagnosis and analysis. Many
methods are used to improve data variety:
5.7.1. Geometric Transformations
It involves the application of elastic deformations, random inversions, and rotations to images, a
method particularly effective in real-world circumstances where orientations may differ dramatically.
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
48
Understanding these technical changes lays a strong basis for the next research on data augmentation
techniques; moreover, it is imperative to be able to "safely" and effectively apply these engineering
improvements. Therefore, it is imperative to assess the "integrity" of the improvement to ensure that the
model achieves better accuracy and efficiency in processing higher data volume. [49].
5.7.2. Cutout and Random Erasing
These techniques involve training images obscuring certain tongue areas of interest. This helps the
system learn to classify tongue images even if some areas are concealed or missing by simulating variances
or partial obstructions that might occur in real-world settings. Studies show that using these techniques helps
tongue image categorization algorithms become more dependable and precise.[50].
5.7.3. Flipping
One finds much higher frequency in flipping the horizontal axis than in the vertical one. Among the
easiest to run, this augmentation has shown success on ImageNet and CIFAR-10. This change does not
maintain labels in MNIST or SVHN datasets for text recognition.[48].
5.7.4. Color space
Usually, digital picture data is expressed as a tensor with dimensions (height × width × color
channels). One quite useful approach is using augmentations in the color channel space. Basic color
augmentations entail separating one color channel, say Red, Green, or Blue. By separating that matrix and
adding two zero matrices from the other color channels, one can quickly convert an image into its
representation in one color channel. Simple matrix operations allow one to easily change the brightness of the
image by means of RGB values. Advanced color augmentations are obtained from image-characterizing
color histograms. Changing the intensity values in these histograms results in different lighting, much like in
photo editing programs. [49].
5.7.5. Cropping
By extracting a center patch from every image, picture cropping is a good processing method for
picture data with different height and width dimensions. Furthermore, arbitrary cropping might have the same
effect as translations. Random cropping and translations differ in that while translations preserve image
spatial dimensions, cropping reduces the input size from (256, 256) to (224, 224). The chosen lowering
threshold for cropping might not be a label-preserving change.[51].
5.7.6. Rotation
Rotation augmentations rotate the image anticlockwise or clockwise around an axis between 1° and
359°.The degree of rotation parameter determines the safety of rotation augmentations in a major part. While
minor rotations between 1 and 20 degrees or −1 and −20 degrees may help with digit recognition tasks like
MNIST, as the degree of rotation increases, the integrity of the data label is damaged following the
transformation.[52].
5.6.7. Translation
Translation of images either horizontally or vertically will help to reduce data positional bias.
Should all of the images in a dataset be centered, the model also has to be tested on exactly centered images.
When the original image is moved in a specific direction, the empty space could be filled with random or
Gaussian noise or a constant number, say 0 or 255. This padding preserves image spatial dimensions after
augmentation. [53].
5.7.8. Noise injection
The process of picture noise injection involves the incorporation of random noise, which is typically
derived from a Gaussian distribution, into the images. The utilization of this technique makes it possible to
replicate different illumination circumstances or to simulate inaccuracies, hence strengthening the capability
of models to gain more resilient characteristics and improving their ability to effectively manage data
variances. Moreno-Barea et al. conducted an evaluation of noise injection on nine datasets from the UCI
collection, and their findings demonstrated that it is effective in enhancing the performance of models. The
incorporation of noise improves the model's capacity to recognize patterns in surroundings that are veiled or
challenging, and as a result, it serves as a crucial asset for enhancing resilience in tasks such as photo
categorization. [54].
5.8. Image preprocessing
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
49
Especially in the study of tongue images, which regularly face difficulties like noise, inconsistent
lighting, and resolution constraints resulting from environmental factors or equipment quality, image
preprocessing is a necessary phase in improving the quality and applicability of photos. Preprocessing
techniques seek to sanitize the accessible data and remove elements that can compromise performance, such
as distorted, fuzzy, or incomplete images. Preprocessing is therefore very important for accurate
interpretation since factors like smoke, poor lighting, or movement during image collection can compromise
the quality of tongue photographs. A basic preprocessing method, grayscale equalization distributes the
intensity values of an image's pixels, enhancing its contrast and brightness. The approach starts with the
study of the histogram of the image, therefore showing the grayscale level distribution. While an uneven
distribution may produce too bright or dark areas, therefore diminishing the visual quality of the image, a
balanced histogram indicates adequate contrast and brightness. Grayscale equalization improves important
characteristics in tongue image processing, including color and texture changes—qualities necessary for
diagnosis. [55].
Image processing is a developing technique whereby important insights are derived from a series of
operations on input images. Usually representing the three main colors—red, green, and blue (RGB), or
grayscale values in monochromatic images, input images are shown as pixels. Two main approaches define
image processing: analogues and digital ones. Whereas digital image processing consists of three basic
phases: preprocessing, augmentation, and presentation, succeeded by knowledge extraction, analogue
processing relates to tangible reproductions, such as printed photographs. Pretreatment methods in digital
image processing are essential to prepare the image for the next analysis: noise reduction, grayscale
normalization, and the elimination of distorted or incomplete images. Essential for medical uses, including
tongue image analysis for diagnosis, these techniques improve image quality and offer more exact feature
extraction and analysis. Combining these techniques helps image processing systems to effectively control
changes in input data and provide consistent, high-quality results.[56].
5.9. Tongue image classification
Especially in traditional and modern medicine, tongue image categorization is regarded as one of the
most interesting uses in medical image analysis. The development of more affordable and powerful
computing systems has attracted increased attention in the study of tongue images for non-invasive diagnosis
and health monitoring. In this discipline, digital image processing is indispensable, as it enables the
extraction of critical information from tongue images to facilitate the identification and classification of
illnesses. The process of tongue picture categorization involves the examination of the tongue's unique
characteristics, such as its color, texture, shape, and coating, in order to identify patterns that are associated
with various health conditions. The features are extracted and processed to produce a numerical
representation of the tongue, which is then compared to a database of known examples for classification. The
precision of tongue image analysis can be influenced by a variety of factors, such as the presence of
impediments (e.g., food detritus or saliva), picture resolution, and illumination. Consequently, it is essential
to implement preprocessing procedures, including noise reduction, grayscale equalization, and picture
enhancement, in order to achieve the highest possible level of performance. [57].
The classification of tongue images typically involves three primary stages: the identification of the
tongue region in the image, feature extraction to determine essential properties, and classification based on
the identified features. In order to improve the system's resilience and accuracy, sophisticated algorithms,
including deep learning models, are frequently implemented. These algorithms are capable of identifying
subtle variations in tongue morphology that may indicate specific health conditions, such as infectious
diseases, diabetes, digestive disorders, or systemic diseases. Tongue image categorization has a wide range of
applications, such as telemedicine, which enables remote health monitoring, and personalized healthcare,
which provides personalized diagnostic insights. By incorporating this technology into diagnostic instruments
and electronic health records, medical evaluations can be rendered more precise and efficient. Additionally,
the field will be further advanced by the implementation of standardized datasets and efficient preprocessing
methods, rendering tongue image classification an essential asset in modern medicine. Figure 10 shows deep
learning techniques for tongue diagnosis analysis. In Figure (a), examples of three different real tongue
images (in terms of color and presence of white areas, and in terms of texture with different depths of fissures
and a relatively thick white layer covering the tongue) are used as inputs for deep learning models. Figure (b)
is a simple schematic of a single ResNet101 deep learning model (CNN) used to classify tongue images. This
single model does all the work (feature extraction and classification). Figure (c) is a system that uses several
completely different deep learning models (e.g. ResNet101, VGG16, DenseNet) together to classify tongue
images. Each model classifies the image independently. The results of the three models are combined by
common fusion methods (e.g. , majority voting, soft voting, and stacking) to obtain the final prediction.[58].
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
50
Figure 10. (a, b, c) view Deep learning methods for tongue diagnosis analyses
6. EVALUATION METRICS
Evaluating Conventional Tools for Tongue Image Classification In tongue image classification,
practical application model performance measurement depends on traditional assessment instruments and
metrics. Analyzing the accuracy, robustness, and efficiency of these models under several evaluation criteria
and benchmarking techniques can help one to evaluate their effectiveness. The main metrics and methods
used for evaluating tongue image categorization systems will be described in this part. The main evaluation
gauges are: [59].
6.1. Accuracy
One of the most important evaluation measures used in many fields, including the analysis of the
tongue photo ratio to the overall total count, is accuracy. High accuracy indicates, as shown in Equation (1),
the model can regularly distinguish between several linguistic elements connected to different health
problems.[60].
Accuracy = !"#$#!%
!"#$#!%#$#&"#$#&% (1)
6.2. Precision
Computes the ratio of precisely expected positive events—that is, tongue images for a certain
condition—to the overall projected positive events. High accuracy means the model produces fewer false
positive mistakes. As Equation (2) illustrates,[61].
Precision = !"
#!"#$#&" (2)
6.3. Recall (Sensitivity)
Evaluates the model's precisely identified real positive case ratio.A high recall indicates that, as
stated in Equation (3), the model ignores fewer real positive cases[62].
Fissured tongue image, tooth-marked tongue image, and tongue image with fissures and tooth marks
CNN method of single-object detection
CNN method of multi-label detection in tongue images
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
51
Recall (Sensitivity) = !"
!"#$#&% (3)
6.4. F1-Score
Considered all-around, including false positives and false negatives, the F1 score is the harmonic
mean of precision and recall and is therefore a vital evaluation statistic for judging the performance of the
model. Particularly helpful for imbalanced datasets and a more consistent assessment of a model's efficacy
than correctness, which may ignore certain mistakes, the F1 Score provides a complete evaluation of model
performance. This results from its evaluation of both mistake types. As shown in Equation (4).[63].
F1 − Score = 2 ×#"'()*+*,-#×#/()011
#"'()*+*,-#$#/()011 (4)
6.5. ROC (Receiver Operating Characteristic)
The ROC curve graphically shows the model's performance at certain classification thresholds.
Plotting the True Positive Rate (TPR) in respect to the False Positive Rate (FPR). This image helps us to find
the moment when the rates of false positives and true positives equal one another. Analyzing the ROC curve
helps one assess the performance of a classification model and guide threshold choosing.[64].
6.6. AUC (Area under the Curve)
For classification problems, the AUC—also known as area under the ROC curve—offers a complete
picture of a model's performance. It produces a single scalar value to evaluate the model's effectiveness.
Whereas an AUC of 0.5 denotes random chance, an AUC of 1 denotes perfect classification. The AUC is
widely used in performance evaluation in many different disciplines and is necessary to assess the
effectiveness of categorization systems. [65].
6.7. Confusion Matrix
Comprising accurate forecasts, erroneous predictions, and false predictions, a confusion matrix
summarizes the model's predictions in respect to the real labels. One can use it to identify areas of concern
and determine several evaluation criteria. True Positives (TP), which are successfully expected positives;
True Negatives (TN), which are correctly predicted negatives; False Positives (FP), which are incorrectly
predicted positives (Type I error); and False Negatives (FN), which are incorrectly predicted negatives (Type
II error).[66].
6.8. FRR (False Rejection Rate)
The False Rejection Rate (FRR) is the percentage of actual positives the model mistakenly rejects. A
low false rejection rate means the model regularly ignores real positive events. As seen in Equation(5).[67].
FRR = &%
#&%#$#!" (5)
6.9. Specificity
Specificity, sometimes known as the true negative rate, measures the model's exact fraction of actual
negatives identified. This statistic clarifies the efficiency in identifying negatives, improving the general
performance and dependability of the system. High specificity is the indication that the model generates
fewer false positive mistakes. As Equation (6) demonstrates.[68].
Speci f icity = !%
!%#$#&" (6)
6.10. True positive rate (TPR)
The ratio of true positives (TP) accurately identified from the total of true positives (TP) and false
negatives (FN) is the recall, sometimes known as the true positive rate. Equation (6) describes how one might
run the computation.[68].
TPR = 2#!"
&%#$#!" (7)
7. Challenges and Future Work
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
52
Future research has many chances to increase the present capabilities of tongue image classification
algorithms to identify diseases and contribute to the future by means of numerous efforts; past research has
faced many obstacles that demand addressing in future research.
7.1. Challenges
1- Imbalanced data: Achieving data balance is difficult since the small differences in picture
amounts across various classes—e.g., healthy tongue versus tongue displaying specific diseases—
are insufficient to classify the dataset as either completely balanced or significantly imbalanced.
When a certain class is much underrepresented, the model may completely ignore that class,
resulting in poor performance and a difficulty to consistently identify rare or uncommon diseases.
2- Tongue image classification:Tongue image classification datasets have to cover not only the
presence of some disorders but also their degree or stage. A tongue with early-stage illness has
different traits than one showing advanced disease. Misclassifying early events as advanced ones
could lead to false diagnosis results, therefore reducing the effectiveness of treatment.
3- Optimizing Model Parameters:Hyperparameter optimization of learning rate, batch size, neural
network layer, and node count directly affects the tongue image classification model's improvement.
Optimal model performance is obtained by means of careful choosing of these parameters, hence
enhancing the model's ability for exact and effective learning.This optimization will directly affect
how well researchers understand the possibilities and limitations of the model.
4- Detecting complex diseases: Some diseases affecting the human body show signs on the tongue,
which can be complex or vague and include subtle changes in color or texture. These changes could
be difficult for systems designed mostly for the detection of common diseases. Therefore, the
accuracy of the model could decrease, which would increase classification mistakes, especially in
cases when the symptoms are vague.
5- Integrating tongue classification with medical diagnostic systems: In modern medical diagnosis
systems, tongue imaging could be the first step; hence, integrating tongue categorization with other
techniques, such as clinical symptom analysis or laboratory tests, can increase the accuracy and
effectiveness of these systems. This method will also assist in addressing the issues related to
misdiagnosis caused by symptom overlap across various diseases.
6- Tongue image accuracy: The precision of tongue images is contingent upon the imaging device's
accuracy, illumination intensity, imaging angle, and the individual's seated posture during imaging,
as well as the tongue's coloration being influenced by the dietary intake of the afflicted individual.
7.2. Future Works
1- Expanding to discover other diseases:Future studies could look at other diseases connected to
the tongue, such as fungal infections, ulcers, and gastrointestinal and hepatic problems, which would
show similar symptoms but call for different treatments.
2- Integration of additional diagnostic methods: Integrating other diagnostic methods, such as voice
analysis to detect speech problems connected to tongue disorders or the examination of alternative
medical images, helps the system's accuracy to be improved.
3- Ethical and Privacy Considerations:Particularly with regard to the preservation of medical
data confidentiality, the ethical and privacy issues of gathering and evaluating tongue pictures must
be addressed as artificial intelligence is used in medical diagnostics more and more.
4- Early Diagnosis Alert System: Establishing a notification system will let users or doctors be
informed upon the discovery of the first signs of tongue problems, therefore promoting quick
diagnosis and efficient treatment.
5- Enhanced security and reliability: To ensure accurate diagnosis and data protection, the system
can be improved to include extra security components, such as confirming the patient's identity
using alternative biometric data.
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
53
8. CONCLUSION
This paper provides a comprehensive analysis of the roles of machine learning and deep learning
techniques in improving disease detection and diagnosis, specifically regarding diabetes and other tongue-
related disorders. It underscores a non-invasive and effective methodology for early disease detection,
highlighting improvements that address issues such as subtle symptom fluctuations, imbalanced datasets, and
practical implementation in real-world contexts.Techniques for data augmentation were employed in order to
improve the diversity of the data and optimize the performance of the model, which finally resulted in
improved symptom detection and disease categorization. A number of cutting-edge technologies, including
convolutional neural networks (CNNs), were applied in order to extract characteristics of the tongue,
including its color, texture, and shape. Deep learning techniques, such as Convolutional Neural Networks
(CNN) and machine learning algorithms, including Support Vector Machines (SVM), Decision Trees (DT),
Logistic Regression (LR), Naïve Bayes, and Random Forests, have shown efficacy in assessing tongue
features like discoloration and texture, thereby improving diagnostic accuracy for diabetes and other
conditions. Research indicates that non-surgical approaches can enhance the cost-effectiveness and comfort
of the diagnostic process. Attaining elevated classification accuracy through the amalgamation of
technologies like support vector machines (SVM) and deep neural networks. The research faced challenges
such as data imbalances and discrepancies in lighting conditions and image quality, which were mitigated
through pre-processing techniques like SMOTE.In the future, research should concentrate on the
development of advanced deep learning and machine learning techniques, as well as the integration of tongue
image analysis with traditional diagnostic tools such as blood tests.The implementation of this will make it
possible to enhance the accuracy and effectiveness of the system in real time while simultaneously giving
emphasis to ethical considerations and safeguarding patient data. In spite of the progress that has been made,
difficulties such as symptom overlap and practical reliability still exist. It is necessary to find solutions to
both specific and ethical problems in order to guarantee the safe handling of data. With the incorporation of
deep learning These instruments make it possible to detect diseases, such as diabetes, at an earlier stage by
the thorough collection of data and the observation of tongue symptoms.
REFERENCES
[1] M. Bharathi, D. Prasad, T. Venkatakrishnamoorthy, and M. Dharani, “Diabetes Diagnostic Method
based on Tongue Image Classification Using Machine Learning Algorithms,” J. Pharm. Negat.
Results, vol. 13, no. 4, pp. 1247–1250, 2022, doi: 10.47750/pnr.2022.13.04.174.
[2] L. T*, G. P, S. V, T. P, and S. M, “Detection of Diabetes Mellitus using Tongue images,” Int. J.
Recent Technol. Eng., vol. 8, no. 4, pp. 3475–3482, 2019, doi: 10.35940/ijrte.c6409.118419.
[3] P. C. Hsu et al., “The tongue features associated with type 2 diabetes mellitus,” Med. (United States),
vol. 98, no. 19, pp. 1–6, 2019, doi: 10.1097/MD.0000000000015567.
[4] J. Zhang et al., “Diagnostic Method of Diabetes Based on Support Vector Machine and Tongue
Images,” Biomed Res. Int., vol. 2017, 2017, doi: 10.1155/2017/7961494.
[5] B. Gabhale, M. Shinde, A. Kamble, and M. Kulloli, “Tongue Image Analysis with Color and Gist
Features for Diabetes Diagnosis.,” Int. Res. J. Eng. Technol., vol. 4, no. 4, pp. 523–526, 2017,
[Online]. Available: https://www.irjet.net/archives/V4/i4/IRJET-V4I4104.pdf
[6] T. Jiang et al., “Tongue image quality assessment based on a deep convolutional neural network,”
BMC Med. Inform. Decis. Mak., vol. 21, no. 1, pp. 1–14, 2021, doi: 10.1186/s12911-021-01508-8.
[7] U. Thirunavukkarasu, S. Umapathy, P. T. Krishnan, and K. Janardanan, “Human Tongue
Thermography Could Be a Prognostic Tool for Prescreening the Type II Diabetes Mellitus,”
Evidence-based Complement. Altern. Med., vol. 2020, 2020, doi: 10.1155/2020/3186208.
[8] A. Stephen Sagayaraj, S. K. Kabilesh, A. Anand Kumar, S. Gokulnath, T. Mani, and K. Dinakaran,
“Diabetes Mellitus and Diabetic Retinopathy Detection using Tongue Images,” J. Phys. Conf. Ser.,
vol. 1831, no. 1, 2021, doi: 10.1088/1742-6596/1831/1/012028.
[9] J. Li et al., “Research of the Distribution of Tongue Features of Diabetic Population Based on
Unsupervised Learning Technology,” Evidence-based Complement. Altern. Med., vol. 2022, 2022,
doi: 10.1155/2022/7684714.
[10] S. Balasubramaniyan, V. Jeyakumar, and D. S. Nachimuthu, “Panoramic tongue imaging and deep
convolutional machine learning model for diabetes diagnosis in humans,” Sci. Rep., vol. 12, no. 1,
pp. 1–18, 2022, doi: 10.1038/s41598-021-03879-4.
[11] B. F. Wee, S. Sivakumar, K. H. Lim, W. K. Wong, and F. H. Juwono, “Diabetes detection based on
machine learning and deep learning approaches,” Multimed. Tools Appl., vol. 83, no. 8, pp. 24153–
24185, 2024, doi: 10.1007/s11042-023-16407-5.
[12] U. Thirunavukkarasu, S. Umapathy, V. Ravi, and T. J. Alahmadi, “Tongue image fusion and analysis
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
54
of thermal and visible images in diabetes mellitus using machine learning techniques,” Sci. Rep., vol.
14, no. 1, pp. 1–17, 2024, doi: 10.1038/s41598-024-64150-0.
[13] B. Tiryaki, K. Torenek-Agirman, O. Miloglu, B. Korkmaz, İ. Y. Ozbek, and E. A. Oral, “Artificial
intelligence in tongue diagnosis: classification of tongue lesions and normal tongue images using
deep convolutional neural network,” BMC Med. Imaging, vol. 24, no. 1, pp. 1–9, 2024, doi:
10.1186/s12880-024-01234-3.
[14] J. K. Mathew and S. Sathyalakshmi, “Sine hunter prey optimization enabled deep residual network
for diabetes mellitus detection using tongue image,” J. Assoc. Med. Sci., vol. 57, no. 2, pp. 76–85,
2024, doi: 10.12982/JAMS.2024.029.
[15] L. Wu, X. Luo, and Y. Xu, “Using convolutional neural network for diabetes mellitus diagnosis
based on tongue images,” J. Eng., vol. 2020, no. 13, pp. 635–638, 2020, doi: 10.1049/joe.2019.1151.
[16] M. M. Taye, “Understanding of Machine Learning with Deep Learning :,” Comput. MDPI, vol. 12,
no. 91, pp. 1–26, 2023.
[17] J. Li et al., “Establishment of noninvasive diabetes risk prediction model based on tongue features
and machine learning techniques,” Int. J. Med. Inform., vol. 149, no. May, 2021, doi:
10.1016/j.ijmedinf.2021.104429.
[18] J. Li et al., “A multi-step approach for tongue image classification in patients with diabetes,”
Comput. Biol. Med., vol. 149, no. October, 2022, doi: 10.1016/j.compbiomed.2022.105935.
[19] S. Fan et al., “Machine learning algorithms in classifying TCM tongue features in diabetes mellitus
and symptoms of gastric disease,” Eur. J. Integr. Med., vol. 43, 2021, doi:
10.1016/j.eujim.2021.101288.
[20] V. H. Phung and E. J. Rhee, “A High-accuracy model average ensemble of convolutional neural
networks for classification of cloud image patches on small datasets,” Appl. Sci., vol. 9, no. 21, 2019,
doi: 10.3390/app9214500.
[21] N. Thomas Rincy and R. Gupta, “Ensemble learning techniques and its efficiency in machine
learning: A survey,” 2nd Int. Conf. Data, Eng. Appl. IDEA 2020, 2020, doi:
10.1109/IDEA49133.2020.9170675.
[22] I. D. Mienye and Y. Sun, “A Survey of Ensemble Learning: Concepts, Algorithms, Applications, and
Prospects,” IEEE Access, vol. 10, no. September, pp. 99129–99149, 2022, doi:
10.1109/ACCESS.2022.3207287.
[23] M. Shah, H. Kantawala, K. Gandhi, R. Patel, K. A. Patel, and A. Kothari, “Theoretical Evaluation of
Ensemble Machine Learning Techniques,” Proc. - 5th Int. Conf. Smart Syst. Inven. Technol. ICSSIT
2023, pp. 829–837, 2023, doi: 10.1109/ICSSIT55814.2023.10061139.
[24] Y. Cao, T. A. Geddes, J. Y. H. Yang, and P. Yang, “Ensemble deep learning in bioinformatics,” Nat.
Mach. Intell., vol. 2, no. 9, pp. 500–508, 2020, doi: 10.1038/s42256-020-0217-y.
[25] A. Kumar and M. Jain, “Ensemble Learning for AI Developers: Learn Bagging, Stacking, and
Boosting Methods with Use Cases,” Ensemble Learn. AI Dev. Learn Bagging, Stacking, Boost.
Methods with Use Cases, no. January, pp. 1–136, 2020, doi: 10.1007/978-1-4842-5940-5.
[26] D. W. Opitz and R. F. MacLin, “An empirical evaluation of bagging and boosting for artificial neural
networks,” IEEE Int. Conf. Neural Networks - Conf. Proc., vol. 3, pp. 1401–1405, 1997, doi:
10.1109/ICNN.1997.613999.
[27] P. Bühlmann, “Bagging, boosting and ensemble methods,” Handb. Comput. Stat. Concepts Methods
Second Ed., vol. 6, no. December, pp. 985–1022, 2012, doi: 10.1007/978-3-642-21551-3__33.
[28] U. Sarmah, P. Borah, and D. K. Bhattacharyya, “Ensemble Learning Methods: An Empirical Study,”
SN Comput. Sci., vol. 5, no. 7, 2024, doi: 10.1007/s42979-024-03252-y.
[29] T. Tong and Z. Li, “Predicting learning achievement using ensemble learning with result
explanation,” PLoS One, vol. 20, no. 1, pp. 1–25, 2025, doi: 10.1371/journal.pone.0312124.
[30] R. O. Odegua, “An Empirical Study of Ensemble Techniques (Bagging, Boosting and Stacking)
Rising Odegua Nossa Data An Empirical Study of Ensemble Techniques (Bagging, Boosting and
Stacking),” Proc. Conf. Deep Learn, no. January, 2019, [Online]. Available:
https://www.researchgate.net/publication/338681864
[31] J. Gui, Z. Sun, Y. Wen, D. Tao, and J. Ye, “A Review on Generative Adversarial Networks:
Algorithms, Theory, and Applications,” IEEE Trans. Knowl. Data Eng., vol. 35, no. 4, pp. 3313–
3332, 2023, doi: 10.1109/TKDE.2021.3130191.
[32] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, “Generative
Adversarial Networks: An Overview,” IEEE Signal Process. Mag., vol. 35, no. 1, pp. 53–65, 2018,
doi: 10.1109/MSP.2017.2765202.
[33] Z. Pan, W. Yu, X. Yi, A. Khan, F. Yuan, and Y. Zheng, “Recent Progress on Generative Adversarial
Networks (GANs): A Survey,” IEEE Access, vol. 7, pp. 36322–36333, 2019, doi:
IJICI ISSN: 2791-2868 r
A comprehensive review of advances in tongue image classification techniques for diabetes identification
(Ghazwan H. Hussien)
55
10.1109/ACCESS.2019.2905015.
[34] L. Malviya, S. Mal, P. Lalwani, and J. S. Chadha, “Diabetes Classification Using Machine Learning
and Deep Learning Models,” Lect. Notes Electr. Eng., vol. 796, pp. 487–503, 2021, doi:
10.1007/978-981-16-5078-9_40.
[35] D. V Mankar and P. S. Chaudhary, “Tongue Image Diagnosis System using Machine Learning with
Hand-Crafted Features,” vol. 7588, no. 6, 2024, doi: 10.54105/ijpmh.L1097.04060924.
[36] I. H. Sarker, “Machine Learning: Algorithms, Real-World Applications and Research Directions,”
SN Comput. Sci., vol. 2, no. 3, pp. 1–21, 2021, doi: 10.1007/s42979-021-00592-x.
[37] C. Anusha, C. Anusha, and C. Engineering, “A Machine Learning Approach for Prediction of
Diabetes Mellitus,” Int. J. Emerg. Trends Eng. Res., vol. 11, no. 6, pp. 207–213, 2023, doi:
10.30534/ijeter/2023/031162023.
[38] K. Kangra and J. Singh, “Comparative analysis of predictive machine learning algorithms for
diabetes mellitus,” Bull. Electr. Eng. Informatics, vol. 12, no. 3, pp. 1728–1737, 2023, doi:
10.11591/eei.v12i3.4412.
[39] M. Banday, S. Zafar, P. Agarwal, and M. Afshar, “Utilising Machine Learning Algorithms for
Predictive Analysis of Diabetes,” 2021.
[40] R. D. Joshi and C. K. Dhakal, “Predicting type 2 diabetes using logistic regression and machine
learning approaches,” Int. J. Environ. Res. Public Health, vol. 18, no. 14, 2021, doi:
10.3390/ijerph18147346.
[41] N. A. Farooqui, . R., and A. Tyagi, “Prediction Model for Diabetes Mellitus Using Machine Learning
Techniques,” Int. J. Comput. Sci. Eng., vol. 6, no. 3, pp. 292–296, 2018, doi:
10.26438/ijcse/v6i3.292296.
[42] S. Jagadeesan, A. Kapoor, and S. Ghosh, “Diabetes Prediction Using Machine Learning,” AIP Conf.
Proc., vol. 3075, no. 1, pp. 294–305, 2024, doi: 10.1063/5.0217181.
[43] M. Chaubey, “Diabetes Mellitus Prediction using Machine Learning,” Int. J. Res. Appl. Sci. Eng.
Technol., vol. 11, no. 5, pp. 4786–4790, 2023, doi: 10.22214/ijraset.2023.52755.
[44] P. Rani, R. Lamba, R. K. Sachdeva, P. Bathla, and A. N. Aledaily, “Diabetes Prediction Using
Machine Learning Classification Algorithms,” Int. Conf. Smart Comput. Appl. ICSCA 2023, pp.
1664–1669, 2023, doi: 10.1109/ICSCA57840.2023.10087827.
[45] Z. Loukil, Q. K. A. Mirza, W. Sayers, and I. Awan, “A Deep Learning based Scalable and Adaptive
Feature Extraction Framework for Medical Images,” Inf. Syst. Front., vol. 26, no. 4, pp. 1279–1305,
2024, doi: 10.1007/s10796-023-10391-9.
[46] S. M. Vijithananda et al., “Feature extraction from MRI ADC images for brain tumor classification
using machine learning techniques,” Biomed. Eng. Online, vol. 21, no. 1, pp. 1–21, 2022, doi:
10.1186/s12938-022-01022-6.
[47] D. Rastogi et al., “Deep learning-integrated MRI brain tumor analysis : feature extraction ,
segmentation , and Survival Prediction using Replicator and volumetric networks,” 2025.
[48] E. D. Cubuk, B. Zoph, D. Mane, V. Vasudevan, and Q. V. Le, “Autoaugment: Learning
augmentation strategies from data,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit.,
vol. 2019-June, no. Section 3, pp. 113–123, 2019, doi: 10.1109/CVPR.2019.00020.
[49] C. Shorten and T. M. Khoshgoftaar, “A survey on Image Data Augmentation for Deep Learning,” J.
Big Data, vol. 6, no. 1, 2019, doi: 10.1186/s40537-019-0197-0.
[50] H. Bagherinezhad, M. Horton, M. Rastegari, and A. Farhadi, “Label Refinery: Improving ImageNet
Classification through Label Progression,” no. 1, 2018, [Online]. Available:
http://arxiv.org/abs/1805.02641
[51] E. Randellini, L. Rigutini, and C. Saccà, “Data Augmentation Techniques and Transfer Learning
Approaches Applied to Facial Expressions Recognition Systems,” Int. J. Artif. Intell. Appl., vol. 13,
no. 1, pp. 55–72, 2022, doi: 10.5121/ijaia.2022.13104.
[52] S. Porcu, A. Floris, and L. Atzori, “Evaluation of data augmentation techniques for facial expression
recognition systems,” Electron., vol. 9, no. 11, pp. 1–12, 2020, doi: 10.3390/electronics9111892.
[53] K. G. Kim and B. T. Lee, “Graph structure based data augmentation method,” Biomed. Eng. Lett.,
2024, doi: 10.1007/s13534-024-00446-4.
[54] J. Sheng, J. Fan, P. Ye, and J. Cao, “JNDMix: Jnd-Based Data Augmentation for No-Reference
Image Quality Assessment,” ICASSP, IEEE Int. Conf. Acoust. Speech Signal Process. - Proc., 2023,
doi: 10.1109/ICASSP49357.2023.10096234.
[55] Y. Sun, “High-resolution image processing and entity recognition algorithm based on artificial
intelligence,” J. Intell. Syst., vol. 33, no. 1, 2024, doi: 10.1515/jisys-2023-0245.
[56] H. K. Gutlapalli, “Research Opportunities and Challenges in Image Processing,” Int. Trans. Electr.
Eng. Comput. Sci., vol. 2, no. 3, pp. 102–111, 2023, doi: 10.62760/iteecs.2.3.2023.57.
r ISSN: 2791-2868
I Vol. 4, 1, June 2025, pp. 33-56
56
[57] Q. Liu et al., “A survey of artificial intelligence in tongue image for disease diagnosis and syndrome
differentiation,” Digit. Heal., vol. 9, 2023, doi: 10.1177/20552076231191044.
[58] T. Jiang et al., “Deep Learning Multi-label Tongue Image Analysis and Its Application in a
Population Undergoing Routine Medical Checkup,” Evidence-based Complement. Altern. Med., vol.
2022, 2022, doi: 10.1155/2022/3384209.
[59] N. Sulayman, “Predicting Type 2 Diabetes Mellitus using Machine Learning Algorithms Digital
Image Processing View project,” vol. 44, no. 2022, pp. 89–100, [Online]. Available:
https://www.researchgate.net/publication/366634353
[60] M. F. Faruque, Asaduzzaman, and I. H. Sarker, “Performance Analysis of Machine Learning
Techniques to Predict Diabetes Mellitus,” 2nd Int. Conf. Electr. Comput. Commun. Eng. ECCE 2019,
2019, doi: 10.1109/ECACE.2019.8679365.
[61] M. T. Islam, M. Raihan, F. Farzana, M. G. M. Raju, and M. B. Hossain, “An Empirical Study on
Diabetes Mellitus Prediction for Typical and Non-Typical Cases using Machine Learning
Approaches,” 2019 10th Int. Conf. Comput. Commun. Netw. Technol. ICCCNT 2019, 2019, doi:
10.1109/ICCCNT45670.2019.8944528.
[62] A. Shitole, A. Kenchappagol, R. Jangle, Y. Shinde, and A. S. Chadha, “Enhancing Retinal Scan
Classification: A Comparative Study of Transfer Learning and Ensemble Techniques,” Int. J. Recent
Innov. Trends Comput. Commun., vol. 11, no. 7s, pp. 520–528, 2023, doi:
10.17762/ijritcc.v11i7s.7031.
[63] Z. N. Abed and A. M. Al-Bakry, “Diagnose Eyes Diseases Using Various Features Extraction
Approaches and Machine Learning Algorithms,” Iraqi J. Comput. Informatics, 2023.
[64] J. S. Akosa, “Predictive Accuracy: A Misleading Performance Measure for Highly Imbalanced
Data,” SAS Glob. Forum, vol. 942, pp. 1–12, 2017, [Online]. Available:
https://support.sas.com/resources/papers/proceedings17/0942-2017.pdf
[65] I. M. De Diego, A. R. Redondo, R. R. Fernández, J. Navarro, and J. M. Moguerza, “General
Performance Score for classification problems,” Appl. Intell., vol. 52, no. 10, pp. 12049–12063,
2022, doi: 10.1007/s10489-021-03041-7.
[66] J. N. de Alencar, M. H. de J. Oliveira, M. C. N. Sampaio, M. F. Rego, and R. Nunes, “A Journey
Through Philosophy and Medicine: From Aristotle to Evidence-Based Decisions,” Philosophies, vol.
9, no. 6, pp. 1–15, 2024, doi: 10.3390/philosophies9060189.
[67] M. H. Murray and J. D. Blume, “False Discovery Rate Computation: Illustrations and
Modifications,” pp. 1–18, 2020, [Online]. Available: http://arxiv.org/abs/2010.04680
[68] J. Terven, D. M. Cordova-Esparza, A. Ramirez-Pedraza, E. A. Chavez-Urbiola, and J. A. Romero-
Gonzalez, “Loss Functions and Metrics in Deep Learning,” 2023, [Online]. Available:
http://arxiv.org/abs/2307.02694
