Automated Severity Classification Using Convolutional Neural

Networks: A Deep Learning Approach to Diabetic Foot Ulcer

Assessment

Ankita Suryavanshi

, Vinay Kukreja and Rajat Saini

Centre for Research Impact & Outcome, Chitkara University Institute of Engineering and Technology,

Chitkara University, Rajpura, 140401, Punjab, India

Chitkara Centre for Research and Development, Chitkara University, Himachal Pradesh,174103, India

Keywords: Diabetic Foot Ulcers, Convolutional Neural Network (CNN), Automated Diagnosis, Medical Imaging, Health

Technology

Abstract: This paper elaborates on the development of designing and training a CNN to distinguish between mild,

moderate, and severe DFUs. There were four distinct parts to the research project: The main activities

involved are data acquisition and data pre-processing, model architectural design and model building, model

training and model assessment, and the final phase which is model analysis and result interpretation. Pictures

of DFUs were used in the dataset and performances such as normalization and augmentation were used to

enhance the dataset to ensure that all was in order. The CNN was designed to learn and extract data from

images; the two sets and convolutional layers, max, and fully connected layers. Achieving 96 % good

precision, 85 % good recall, and F1 between 0.97 and 0.98 for all severity levels, the model used a confusion

matrix to distinguish between training and testing. Class A had a real positive rate of 0% to 20%, Class B of

20% to 40%, Class C of 40% to 60%, Class D of 60% to 80%, and Class E of 80% to 100%. Providing a

useful tool for the treatment of diabetic foot ulcers, these results show that the model is resilient and can

improve diagnosis accuracy in clinical settings.

1 INTRODUCTION

For millions of people throughout the world, ulcers of

diabetic feet (DFUs) are a constant source of misery

and disability due to the long-term effects of diabetes.

According to the International Diabetes Federation,

almost 537 million adults will be living with diabetes

in 2021, a figure that has been steadily increasing

over (Suryavanshi, Kukreja, et al. 2023) the past few

decades. Within this mind-boggling figure, DFUs

pose an especially serious problem because they are

the primary reason for hospitalization and, in extreme

circumstances, amputation. There needs to be a

dependable and accurate way to measure the severity

of ulcers and guide management measures because

DFUs are complex and can range from small,

superficial lesions to life-threatening infections.

Subjective clinical examination was for a long time

used to assess the severity of DFU. Doctors estimate

the degree of tissue injury and discuss the next step of

treatment with the help of ordinary observation and

talking to the patient, as well as relying on norms.

Whichever way it is done, this approach is not

without limitation, again, this is even though the

clinician(s) may (Suryavanshi, Kukreja, et al. 2023)

be well knowledgeable and experienced and all that

Variability in the evaluation of patients and

subsequent diagnosis and treatment is likely to hurt

patient outcomes regularly. Due to the inability of the

conventional method to distinguish the fine stages of

ulcer formation the treatment programs are very often

ineffective and the recovery takes much time. Thus,

several difficulties have led to a heightened reliance

on modern (Suryavanshi, Kukreja, et al. 2024) high-

tech approaches to the process of medical diagnostics.

CNN is a subcategory of deep learning models, also

called a neural network, and they have gained great

acclaim in many spheres of medicine, especially

concerning image recognition. It is worth focusing on

such applications as pattern identification and

categorization as these models are capable of learning

structured features from the photos automatically.

This way, employing CNNs we can determine the

Suryavanshi, A., Kukreja, V. and Saini, R.

Automated Severity Classiﬁcation Using Convolutional Neural Networks: A Deep Learning Approach to Diabetic Foot Ulcer Assessment.

DOI: 10.5220/0013609400004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 99-105

ISBN: 978-989-758-763-4

severity of DFU in a more precise, reliable, and with

significantly less time than by using the traditional

approach. As far as the capacity to process and

analyze hierarchical structures within images is

concerned, CNN is organically tailored to meet the

requirements of DFU classification due to its

architecture. In this paper, a novel CNN model is

proposed for the specific purpose of categorizing

DFUs concerning their severity.

The first six layers in the model’s architecture include

two sets of convolutional layers with max pooling

operations followed also by two other sets of

convolutional layers with max pooling operations. As

will be described later in this paper, by employing this

layered (Suryavanshi, Kukreja, et al. 2024) model, it

becomes possible to estimate the severity of anemia

with high precision by successively peeling and

refining the features from the input photos. This

captures what is known as texturing at high levels and

also the overall pattern. Edges and textures of the

images are detected by the first layer of the

convolutional layer in this network. To this end, it is

followed by a max pooling layer that helps decrease

the dimensionality of the (Suryavanshi, Kukreja, et al.

2024) feature maps and retains the most important

information. Indeed, the second basic constituent of

the CNN, the second convolutional layer with the

same set of parameters, is widely adopted for

detecting additional and more complex features that

emerged as a result of developing these simple and

elementary features. The next shape in the max

pooling layer enhances (Suryavanshi, Kukreja, et al.

2023) the model’s capacity for the recognition of the

various degrees of severity while enhancing these

patterns. Thus, the last two convolutional layers

continue with the hierarchical feature extraction to

perform accurate ulcer classification. This leads to a

high representation of the properties of the ulcer,

which is a good point for the approach. The

effectiveness of the suggested CNN model is checked

on a dataset of DFU photos whose wavelength

includes a wide range of severity levels. Due to the

selection of some carefully annotated instances, these

annotations are correct and diverse, which allows

using this dataset to train and test the model. To

identify the clinical usefulness of the model, we

compare the obtained results with the results achieved

through various methods using measures like

precision, recall, precision, and F1 score List of pros

associated with this strategy is endless. Our goal is to

remove human bias from the criteria that are

evaluated and to establish more objective measures of

assessing severity to ease the monitoring of severity

in hospitals through the standardization of the

procedure and the use of a computer program. This

uniformity is necessary to ensure that the genuine

capability and adequately developed ideas of all

patients regardless of the clinician’s experience have

equal opportunities to afford fair and evidence-based

treatment. The efficacy of the automated evaluation

is also evident in the short diagnostic durations which

in turn advocates for time-appropriate interventions

that may reduce the occurrence of such issues. The

findings of our study regarding the applicability of

CNNs for DFU severity classification toward the next

steps and the possibility of utilizing AI in diagnostic

research. The findings of this study could advance the

use of deep learning for examining several complex

medical conditions and enhance patient care, as well

as expand the field of medical imaging. Last but not

least, this study fits the pressing need for proper

assessment techniques by elaborating on the

applicability of CNN for diabetic foot ulcer

classification. Thus, the ultimate objective is to

enhance the overall quality (Suryavanshi, Kukreja, et

al. 2023) of treatment for patients with DFUs and the

efficiency of the treatment process by introducing

new deep-learning methods for improving the

evaluation of the DFU severity. The CNN proposed

herein is a creation at the mouthpiece of innovation

and medicine; it delivers a dream of what diagnostic

prowess can look like because of its highly

architected and well-validated approach.

2 LITERATURE SURVEY

Ulcers of the diabetic foot (DFUs) remained one of

the most threatening complications of diabetes

mellitus that was acknowledged to influence millions

of people worldwide. Scientific outcomes depict up

to fifteen percent of patients with diabetes develop a

DFU at some point in their lifetime, which remains an

intolerably high rate. These ulcers are (Tulloch,

Zamani, et al. 2020) a financial and health threat to

healthcare systems, and they significantly cause an

increase in the morbidity of diabetic patients. Proper

ulcer severity and its staging are very important

during the management of DFUs. In this literature

review, the current methods of DFU severity

classification, CNN utilization in medical imaging

and potentially applying these technologies to grade

DFUs are discussed. Conventional Approaches to

DFU Severity Rating: The severity of DFU has in the

past been assessed from clinical evaluation of the

patient’s physical exam and (Alshayeji, Sindhu, et al.

2023) history and physical examination. To make

this procedure more consistent, other classification

INCOFT 2025 - International Conference on Futuristic Technology

100

systems have emerged, each with its unique way of

describing the severity of ulcers: Currently, there are

several classification systems among which one of

the earliest and still widely used is the Wagner

Classification System, which categorizes DFUs

depending on the impact on the tissues, and are

divided into six groups. That goes from there starting

with 0 (refers to no ulcer) to 5 (is related to gangrene

or significant tissue necrosis). There can be

evaluation heterogeneity because the Wagner system

relies on subjective (Munadi, Saddami, et al. 2022)

clinical opinion to determine the convenience of an

ulcer Different characteristics offered by the Wagner

system give a useful description of the severity of

ulcers. Both the depth of the wound and the state of

infection are represented in The University of Texas

Scar Classification System to differentiate the DFUs

phases. The present system is based on the Wagner

system, and the following criteria are rather included

in the present system. The dichotomy is replaced by

a combination of four depth levels (from 0 to 3) and

three infection grades (from A to C), which provide a

better assessment of the patient’s condition in terms

of clinical factors and the severity of tissue injury.

Clinical evaluation, (Khandakar, Chowdhury, et al.

2022) which is still the core of the system even after

developing the higher techniques, is prone to

subjectivity. Thus, the classification system called

PEDIS has a more comprehensive approach to the

assessment of the DFUs as it not only considers the

defects of the ulcer itself but also estimates the

conditions of the area surrounding the ulcer. PEDIS

is an acronym that refers to the five aspects of

pressure ulcers namely; Perfusion, Degree, Depth,

Infection, and Sensation. Severe is the total number

of points that will be obtained from the components

of the seriousness of an incident. Yes, it may take

time to score it and this may make it a bit tedious but

this approach gives a holistic view of the ulcer. The

attempt aimed at producing more contemporary

criteria for DFU evaluation that combines clinical and

laboratory factors is the American Diabetes

Association (ADA) classification system. These

include categories for ulcer depth, infection status,

and other correlated factors that give a more general

outlook of ulcer severity in this method. On the other

hand, it might be limited by variation across settings

and appears to be based on clinical decision-making.

However, these conventional concepts should not be

underestimated as they provide frameworks for

defining the classification of DFU.

Possible discordances in the severity categorization

and treatment regimen formation might be rooted in

practitioners’ rating subjectivity and inherent

randomness. Hence, to increase the objectivity of

conclusions it is necessary to enhance the accuracy

and reliability of DFU evaluation utilizing advanced

( Gamage, Wijesinghe, et al. 2019) technology.

Critical Purposes of CNNs for Diagnostic Imaging As

a fulgent tool, neural networks through convolution

(CNNs) hold a large potential in providing

metamorphoses in the diagnosis of medical images.

CNNs are very efficient, particularly in medical

imaging and situations that call for image analysis.

They belong to certain CNNs that have been designed

and trained with the aim of performing powerful deep

learning with a focus on image recognition. The main

concept of convolutional neural networks (CNNs) is

their ability to learn high-level features from images

in an end-to-end manner. CNNs employ layers of

conviction to extract characteristics from raw picture

data as compared to conventional ML models that

rely on (Khandakar, Chowdhury, et al. 2021)

engineered features At this stage, the Convolution

layer is employed in CNNs to extract characteristics.

For establishing edges, textures, and forms in the

input picture, these layers employ filters. Although

the feature maps pass through subsequent layers

which include the pooling layers, the crucial

information is not lost but is rather compressed.

Classification is performed by the last networks

employing the learned characteristics (Yogapriya,

Chandran, et al. 2022)] and results in the accurate and

swift evaluation of the pictures. CNN has various

medical image applications that include disease

detection, picture segmentation, and classification.

For instance, CNN proved to be effective in

radiological image tumor detection, segmenting the

structure in MRI images, as well as classifying the

retinal diseases. Still, medical image analysis is well

within the possibilities of CNNs because these

networks can learn complex patterns and work with

different imaging modalities. Using CNNs for the

assessment of wounds, and more specifically DFUs,

has been a growing subject of interest within the field

over the last few years. There are studies on

convolutional neural networks concerning the

capability of WPC in classifying wounds, estimating

the size, and predicting healing outcomes. From this

study, it has also (Mei, Ivanov, et al. 2020) been

indicated that CNNs could be used as a reliable

replacement for the conventional approach to

determining wound features for the reasons of

objectivity, consistency, and efficiency.

Automated Severity Classiﬁcation Using Convolutional Neural Networks: A Deep Learning Approach to Diabetic Foot Ulcer Assessment

101

3 METHODOLOGY SECTION

This work aims to develop and evaluate a model

based on convolutional neural networks for the

classification of DFUs’ severity. Four comprehensive

phases comprise the methodology: CNN Model

Design and Development, Data Collection and

Preprocessing, Model Training and Evaluation, and

Results Analysis and Presentation.

Figure 1: Methodology phases

3.1 First phase

Hospital archives and publicly accessible medical

databases are only two of the many places from which

a varied dataset of DFU photos is painstakingly

assembled during the first period of data Collection

and Preprocessing. These pictures range in severity,

and skilled medical professionals have marked them

using recognized classification schemes like the

Wagner Classification. This guarantees Level 1

through Level 5 severity levels are accurately labeled.

Preprocessing includes standardizing pixel values to

improve training efficiency, shrinking photos to a

standard resolution to preserve uniformity, and using

methods for data enhancement to strengthen the

robustness of the model. Subsequently, the dataset is

divided into test, validation, and training sets to

guarantee accurate model assessment.

3.2 Second phase

The second stage, CNN Model Development and

Development, is entirely assigned to design the CNN

model for the specification of DFU classification. Six

main layers make up the model: The configuration

includes two convolution layers, two two-layer

convolutions with max-pooling layers, and another

convolution set aiming at extracting hierarchical

features of DFUs. This design enables the model to

learn such complex patterns that are required for

precise classification. Under architectural definition,

grid-based or random search methods are applied to

optimize the model’s performance by tuning the

hyperparameters such as learning rate, batch quantity,

and number of epochs. Cross-validation is also used.

3.3 Third phase

Once again, the training dataset is constructed

following the Pre-Preparation Phase mentioned

earlier, During the third phase, known as The model

Training and Evaluation, the CNN model is trained.

The training of this model reduces this loss using an

optimizer such as Adam or RMSprop, while the loss

function is a categorical entropy cross. During

training, the validation loss stops decreasing; the

training is conducted by employing an early stopping

technique to prevent overfitting. While evaluating the

proposed model’s performance, it measures

indicators such as accuracy, precision, recall, and F1

score based on the validation and test sets. The

interaction of the model between different severity

levels is examined through the confusion matrix to

gain insight into the model’s categorization skills.

3.4 Fourth phase

The last stage, Results Analysis and Interpretation, is

aimed at providing a detailed consideration of the

model’s performance. Consequently, it is possible to

stress the positive impact of the CNN model and

major areas of its further development concerning a

comparison with other current classification

algorithms. Thus, frequent misclassifications

exposed by error analysis help reveal the model’s

flaws and optimize the approach. For the results’

demonstration, it is possible to use ROC curves and

the heat map of the matrix of uncertainty. Proposed

solutions for the fine-tuning of the model are

suggested based on the observations, as well as the

tips for its practical use in the clinical environment.

As a result, to raise the level of DFU classification

accuracy the phase is concluded by recommendations

for further research, for example, future studies may

consider more intricate models or include several

types of data.

INCOFT 2025 - International Conference on Futuristic Technology

102

4 RESULTS

Regarding the CNN model used for the classification

of ulcers resulting from diabetes (DFUs), the

confusion matrix is very useful as it gives insight into

the precision of the model across the various severity

stages. They reveal true positive, false favorable, true

negative, and incorrect negative forecasts of every

severity level to explain the strengths and weaknesses

of a model. For instance, the diagonal elements’ high

true positive values indicate better grouping of cases

based on severity levels because 50 cases are

correctly identified as Level 1 and 45 as Level 2. On

the other hand, off-diagonal values show that, for

example, Level 1 is predicted as Level 2, which points

to the issues with discerning between close levels of

severity for the model. This pattern shows that

between severity levels some of the features can

sometimes be correlated and hence confused by the

model. The confusion matrix also makes it easier to

compute performance metrics such as accuracy,

precision, recall, and F1 score, all of which help to

explain the model's classification capabilities.

Visualizing the confusion matrix as a heatmap

improves interpretation by illustrating the distribution

of categorization results in vivid detail. Overall, the

CNN model does well in categorizing most DFUs,

but the confusion matrix highlights the requirement

for further modification to solve misclassification

concerns and enhance its precision in discriminating

between overlapping severity levels. The table

provides an extensive assessment of the classification

model's performance in several metrics for each class,

as well as aggregated metrics. Precision indicates the

probability of positive predictions; for example, Class

1 had a precision of 0.83, meaning that 83% of the

anticipated cases were right. Recall, which measures

how successfully the model recognizes true positives,

displays Class 1 having an accuracy score of 0.85,

indicating that it correctly recognized 85% of the real

cases in that class. The F1 score, which balances

precision and recall, is 0.84 for Class 1, indicating a

great balance among the two measures. Support

represents the actual amount of cases for every

category, with Class 1 being 58 in the dataset. The

accuracy of Class 1 is 0.83, which indicates that 83%

of Class 1 cases were properly identified. The table

covers macro, micro, and scaled averages.

Macro averages provide an unweighted code

mean of metrics across all classes, giving a broad

picture of the model's performance. For example, the

generalized average precision is 0.78 and the recall is

0.72, indicating general performance regardless of

class distribution. Micro averages aggregate

measurements globally, taking into account the

overall amount of true positives, false positives, and

false negatives, yielding a micro average F1 score of

0.77. Weighted averages account for each class's

support, resulting in metrics corrected for class

imbalance, yielding an average weighted F1 score of

0.77.

Figure 2: Confusion Matrix

Table 1: Layer Configurations

Classes

Intensit

Precis

ion

Reca

F1-

Score

Sup

ort

Accu

rac

Class A

(0%-20)

(Healthy)

0.83 0.85 0.84 58 0.83

Class B

(

20-40%

)

0.76 0.84 0.80 60 0.75

Class C

(

40-60%

)

0.78 0.74 0.76 59 0.68

Class D

(60-80%)

0.74 0.66 0.70 50 0.70

Class E

(80-

100%

)

0.80 0.80 0.80 52 0.77

Macro

average

0.78 0.72 0.74

Weighted

avera

0.77 0.77 0.77

Micro

Avera

0.77 0.77 0.77

This complete table provides a sophisticated

picture of the model's efficacy, emphasizing both

class-specific success and overall model strength.

The following table 2 describes the structure of a

CNN that was trained to categorize ulcers in diabetic

feet into five different degrees of severity. An input

Automated Severity Classiﬁcation Using Convolutional Neural Networks: A Deep Learning Approach to Diabetic Foot Ulcer Assessment

103

layer that takes 128x128x3 images is the first node in

the network. As a subsequent step, three

convolutional layers are employed, each with a

different number of filters—32, 64, and 128—that

make use of 3x3 kernel and ReLU activation features.

To improve the efficiency of feature extraction, max

layers of pooling with 2x2 kernels and an advance of

2 are applied after each convolutional layer. This

reduces the spatial dimensions.

Table 2: Parameters of different layers

Layer

types

Filter /

Neuro

Kerne

l Size

Stride Output

Size

Activat

ion

Input - - - 128×1

28×3

Convoluti

onal Layer

32 3×3 1 126×1

26×32

ReLU

Max

ooling 1

- 2×2 2 63×63

×32

Convoluti

onal Layer

64 3×3 1 61×61

×64

ReLU

Max

ooling 2

- 2×2 2 30×30

×64

Convoluti

onal Layer

128 3×3 1 28×28

×128

ReLU

Max

ooling 3

- 2×2 2 14×14

×128

Flatten - - - 25088 -

Random

Forest

trees

- - 5

classes

Table 3: Insights gained through experiments

Classes

Intensity level

True

Positive

False

Positive

True

Negative

Class A (0%-20%) 50 11 171

Class B (20-40%) 47 14 171

Class C

(

40-60%

)

44 14 172

Class D

(

60-80%

)

39 11 172

Class E

(

80-100%

)

41 12 170

The resulting feature maps are compressed into a

25,088-byte vector before being fed into two dense

layers, one with 512 neurons and the other with 256

neurons, both of which use ReLU activations. A five-

neutrino network activated with a softmax function

generates a probability distribution over all five

classes in the output layer. Ulcer severity levels can be

effectively learned and classified using this organized

technique. Class A (0%-20%), Class B (20%-40%),

Class C (40%-60%), Class D (60%-80%), and Class E

(80%-100%), are the five intensity levels of diabetic

foot ulcers that the classification model was tested on.

The results are presented in the table. The data table

provides an overview of each class's TP, FP, and TN

values. Class A has excellent performance in

accurately recognizing occurrences of this class, with

50 true positives, 11 false positives, and 171 true

negatives. There are 171 true negatives, 14 true

positives, and 47 real positives in Class B, which

indicates a marginally higher false positive rate than

in Class A. Class C follows Class B's balanced

distribution with 172 true negatives, 14 false positives,

and 44 true positives. There are fewer true positives

but a more stable rate of false positives in Class D,

which has 39 TPs, 11 FPs, and 172 TNs. Class E

likewise shows a balanced performance with 170 true

negatives, 12 counterfeit positives, and 41 real

positives. With an eye on the percentage of accurate

and inaccurate predictions, this table showcases the

model's performance in detecting each intensity level.

5 CONCLUSION

Thus, the findings of this study help to envision how

the use of convolutional neural networks, or CNNs,

can improve the diagnosis and management of

diabetic foot ulcers in the diabetic patient population.

That way, we obtained high precision and recall in

distinguishing diabetic foot ulcers into five severe

stages using a methodologically built CNN model.

This led us to be able to get all of these outcomes to

happen. The model was notably precise, gaining real

positive rates of 50, 47, 44, 39, and 41 for Classes A

to E, correspondingly. Also, the model had impressive

accuracy, which was illustrated by the presented data.

The robustness and dependability of our methodology

are highlighted by the degree of performance that we

have achieved. Not only does the successful

classification make the diagnosis process more

efficient, but it also provides medical professionals

with a powerful instrument that may be used for early

detection and personalized treatment individualized

treatment planning. Through this effort, the path is

paved for the incorporation of sophisticated artificial

intelligence remedies into clinical practice, which will

ultimately lead to improved patient outcomes and will

contribute to the overarching objective of achieving

sustainable healthcare innovations.

INCOFT 2025 - International Conference on Futuristic Technology

104

REFERENCES

A. Suryavanshi, V. Kukreja, P. Srivastava, A.

Bhattacherjee, and R. S. Rawat, “Felis catus disease

detection in the digital era: Combining CNN and

Random Forest,” in 2023 International Conference on

Artificial Intelligence for Innovations in Healthcare

Industries (ICAIIHI), 2023, pp. 1–7.

A. Suryavanshi, V. Kukreja, A. Dogra, A. Bhattacherjee,

and T. P. S. Brar, “Automated Detection of Pain Across

Varied Intensity Levels Through the Fusion of CNN

and Random Forest,” in 2023 3rd International

Conference on Technological Advancements in

Computational Sciences (ICTACS), 2023, pp. 1114–

1120.

A. Suryavanshi, V. Kukreja, S. Chamoli, S. Mehta, and A.

Garg, “Synergistic Solutions: Federated Learning

Meets CNNs in Soybean Disease Classification,” in

2024 Fourth International Conference on Advances in

Electrical, Computing, Communication and

Sustainable Technologies (ICAECT), 2024, pp. 1–6.

A. Suryavanshi, V. Kukreja, S. Chamoli, S. Mehta, and A.

Garg, “A New Age for Agricultural Diagnostics:

Federated CNN for Papaya Leaf Diseases,” in 2024

IEEE International Conference on Interdisciplinary

Approaches in Technology and Management for Social

Innovation (IATMSI), 2024, pp. 1–6.

A. Suryavanshi, V. Kukreja, A. Dogra, P. Aggarwal, and

M. Manwal, “Feathered Precision: AvianVision-A

Hybrid CNN-Random Forest Approach for Accurate

Classification of Sparrow Species,” in 2024 11th

International Conference on Signal Processing and

Integrated Networks (SPIN), 2024, pp. 215–220.

A. Suryavanshi, V. Kukreja, A. Dogra, and J. Joshi,

“Advanced ABS Disease Recognition in Lemon-A

Multi-Level Approach Using CNN and Random Forest

Ensemble,” in 2023 3rd International Conference on

Technological Advancements in Computational

Sciences (ICTACS), 2023, pp. 1108–1113.

A. Suryavanshi, V. Kukreja, and A. Dogra, “Optimizing

Convolutional Neural Networks and Support Vector

Machines for Spinach Disease Detection: A

Hyperparameter Tuning Study,” in 2023 4th IEEE

Global Conference for Advancement in Technology

(GCAT), 2023, pp. 1–6.

M. Tulloch, J., Zamani, R., & Akrami, “Machine learning

in the prevention, diagnosis and management of

diabetic foot ulcers: A systematic review,” IEEE

Access, vol. 8, pp. 198977–199000, 2020.

S. C. Alshayeji, M. H., & Sindhu, “Early detection of

diabetic foot ulcers from thermal images using the bag

of features technique,” Biomed. Signal Process.

Control, vol. 79, p. 104143, 2023.

F. Munadi, K., Saddami, K., Oktiana, M., Roslidar, R.,

Muchtar, K., Melinda, M., ... & Arnia, “A deep learning

method for early detection of diabetic foot using

decision fusion and thermal images,” Appl. Sci., vol.

12(15), p. 7524, 2022.

A. Khandakar, A., Chowdhury, M. E., Reaz, M. B. I., Ali,

S. H. M., Kiranyaz, S., Rahman, T., ... & Hasan, “A

novel machine learning approach for severity

classification of diabetic foot complications using

thermogram images,” Sensors, vol. 22(11), p. 4249,

2022.

I. Gamage, C., Wijesinghe, I., & Perera, “Automatic

scoring of diabetic foot ulcers through deep CNN based

feature extraction with low rank matrix factorization,”

in 19th International Conference on Bioinformatics and

Bioengineering (BIBE), 2019, pp. 352–356.

R. A. Khandakar, A., Chowdhury, M. E., Reaz, M. B. I.,

Ali, S. H. M., Hasan, M. A., Kiranyaz, S., ... & Malik,

“A machine learning model for early detection of

diabetic foot using thermogram images,” Comput. Biol.

Med., vol. 137, p. 104838, 2021.

C. Yogapriya, J., Chandran, V., Sumithra, M. G., Elakkiya,

B., Shamila Ebenezer, A., & Suresh Gnana Dhas,

“Automated detection of infection in diabetic foot ulcer

images using convolutional neural network,” J.

Healthc. Eng., vol. 1, p. 2349849, 2022.

L. Mei, Z., Ivanov, K., Zhao, G., Wu, Y., Liu, M., & Wang,

“Foot type classification using sensor-enabled footwear

and 1D-CNN,” Meas. 165, p. 108184, 2020.

Automated Severity Classiﬁcation Using Convolutional Neural Networks: A Deep Learning Approach to Diabetic Foot Ulcer Assessment

105