Received 27 October 2023, accepted 7 November 2023, date of publication 13 November 2023,

date of current version 17 November 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3332292

Automated Diabetic Foot Ulcer Detection and

Classification Using Deep Learning
SUNNAM NAGARAJU1 , KOLLATI VIJAYA KUMAR 2 , B. PRAMEELA RANI3 , E. LAXMI LYDIA 4,

MOHAMAD KHAIRI ISHAK5,6 , IMEN FILALI7 , FATEN KHALID KARIM 7 ,

AND SAMIH M. MOSTAFA 8
1 Department of Mechanical Engineering, MLR Institute of Technology, Hyderabad 500043, India
2 Department of Computer Science and Engineering, GITAM School of Technology, Vishakhapatnam Campus, GITAM (Deemed to be a University),
Visakhapatnam 530045, India
3 Department of CSE-AIML, Aditya College of Engineering, Surampalem, Andhra Pradesh 533437, India
4 Department of Computer Science and Engineering, Vignan’s Institute of Information Technology, Visakhapatnam 530049, India
5 Department of Electrical and Computer Engineering, College of Engineering and Information Technology, Ajman University, Ajman, United Arab Emirates
6 School of Electrical and Electronic Engineering, Universiti Sains Malaysia, Engineering Campus, Nibong Tebal, Pulau Pinang 14300, Malaysia
7 Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, Riyadh 11671, Saudi Arabia
8 Computer Science Department, Faculty of Computers and Information, South Valley University, Qena 83523, Egypt

Corresponding author: Samih M. Mostafa ([email protected])

This work was supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2023R300),
Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

ABSTRACT Diabetic foot ulcers (DFU) are a common and serious complication in individuals with diabetes,
and early detection plays a crucial role in effective treatment and prevention of further complications.
Automated DFU Detection and Classification using Deep learning (DL) refers to the application of deep
learning techniques to automatically detect and classify diabetic foot ulcers from medical images. DL,
a subfield of machine learning, has shown promising results in medical imaging analysis, including diabetic
foot ulcer detection. The use of deep learning in DFU detection provides various benefits, including the
ability to learn complex features, adaptability to different image modalities, and the potential for high
accuracy in detection and classification tasks. Therefore, this article introduces a novel sparrow search
optimization (SSO) with deep learning enabled diabetic foot ulcer detection and classification (SSODL-
DFUDC) technique. The presented SSODL-DFUDC technique’s goal lies in identifying and classifying
DFU. The proposed technique employs the Inception-ResNet-v2 model for feature vector generation to
accomplish this. Since the trial and error manual hyperparameter tuning of the Inception-ResNet-v2 model is
a tedious and erroneous process, the SSO algorithm can be used for the optimal hyperparameter selection of
the Inception-ResNet-v2 model which in turn enhances the overall DFU classification results. Moreover, the
classification of DFU takes place using the stacked sparse autoencoder (SSAE) model. The comprehensive
experimental outcomes demonstrate the improved performance of the SSODL-DFUDC system related to
existing DL techniques.

INDEX TERMS Medical image analysis, deep learning, diabetic foot ulcer, sparrow search optimization,
computer-aided diagnosis.

I. INTRODUCTION livelihood of quality and distresses social participation and

Amputation of the limb or foot may be caused by a diabetic the outcome of these causes and tissue death because of dis-
foot ulcer (DFU) infection. The probability of survival can be eases. The DFU is increasing rapidly. Because of the scarcity
lesser for patients having amputated limbs [1]. It damages the of specialists and lack of resources in the medication for dia-
betic foot ulcers, over a million diabetic patients at higher risk
The associate editor coordinating the review of this manuscript and of diabetic Mellitus will lose out foot (partly) every year [2].
approving it for publication was Bijoy Chand Chatterjee . Notably, for every 20 seconds, one diabetic foot functioned.
2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
127578 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

A complete study of medical data was essential for spe- ysis of its efficiency. This method utilizes a cross-platform
cialists to accomplish a precise outcome [3]. Conventional mobile structure which allows the utilization of mobile appli-
diagnostic techniques were labour-intensive and vulnerable cations for several environments utilizing a single TypeScript
to human errors. The computer-based diagnostic process’s code base. A DCNN has been used in a cloud-based envi-
utility includes minimal performance enhancement [4]. Cur- ronment, but the mobile application sends photographs of
rent advancements in n wearable and mobile health gadgets patients’ feet to inference for detecting the occurrence of
help in controlling diabetes and its limitations, improving the DFUs. Munadi et al. [15] present a new structure for the
standards of life and extending remission for patients by s DFU classifier dependent upon thermal imaging utilizing
controlling harmful sensing and foot pressure and inflam- DNNs and decision fusion. At this point, decision fusion
mation [5]. The development of novel-generation medical integrates the classifier outcome in a parallel classification.
sensors recommends expanding such devices’ utility in the The author utilized the CNN technique of MobileNetV2
medical field [6]. In the contemporary healthcare mechanism, and ShuffleNet as the baseline classification. In evolv-
medical images are used for diagnosing several patient diffi- ing the classification process, initially, the ShuffleNet and
culties [7]. MobileNetV2 can be trained to utilize plantar thermogram
Conventional approaches for analysis of DFU use databases.
hand-crafted rendered technique [8]. However, research Alshayeji and Sindhu’s [16] investigation is new since,
activities in the publication have displayed that learned without any prior analysis utilized typical ML techniques for
attributes by deep neural networks (DNN) have high poten- the analysis of DFU in the thermal image and also did not
tial compared to classical hand-crafted features [9]. Wide utilize SURF, BOF, or SIFT approaches. Also, the authors uti-
research was conducted to enhance the outcome of com- lize direct temperature files of all feet and then map them on
puterized are very popular in this sector since they were images to obtain the correct temperature distribution. Typical
superior to other techniques [10]. The common method that is ML techniques are chosen in the last step to binary classifier
utilized in the DL technique in medical image classification betwixt normal and DFU. D’Angelo et al. [17] examine a
is convolutional neural networks (CNNs). The CNNs can method dependent upon Genetic Programming (GP) for cre-
efficiently derive valuable attributes for image segmentation, ating an easy global explainable classifier termed X-GPC that
image classification, other vision tasks, and object detec- different present tools, like SHAP and LIME, offer a global
tion [11]. With the obtainability of large-scale trained data analysis of DFU with the mathematical process. Moreover,
and high-performing modern application specific integrated the medicinal clinical staff was offered a simple consultable
circuit (ASICs) and graphics processing units (GPUs) and 3D graph that is utilised to understand the patient’s condition
techniques related to CNNs have enhanced the precision of and make decisions for patient recovery. In [18], a new Deep
image classification. Renowned CNNs for general image CNN, DFU_QUTNet, is also presented for the automatic
classifier tasks involve ResNet, AlexNet, EfficientNet, and classifier of normal skin (healthy skin) towards abnormal skin
VGG [12]. Such networks generally serve as the pillar of (DFU) classes. Stacking further layers to a classic CNN to
medical image classifier networks or directly enforce med- arrive very deep could not cause optimum efficacy, resulting
ical image classifications by transfer learning (TL) with in worse execution because of the gradient.
pre-trained variables on large-scale datasets (e.g., ImageNet). Goyal [19] progressed automatic computer vision (CV)
Practically, labelling and collecting medical images were schemes which recognize the DFU of distinct grades and
costly [13]. TL was a potential way to overcome the need steps. Primarily, the authors utilized ML techniques for clas-
for medicinal trained data. sifying the DFU spots against normal skin spots of the foot
This article introduces a novel sparrow search optimization area to determine the feasible misclassified reasons for both
with deep learning enabled diabetic foot ulcer detection and classes. Secondarily, the authors utilized FCN for the seg-
classification (SSODL-DFUDC) technique. The goal of this mentation of DFU and neighbouring skin from entire foot
technique lies in the identification and classification of DFU. images. Lastly, the authors utilized robust and lightweight
To perform this task, the SSODL-DFUDC technique employs deep localization schemes from mobile devices for detect-
the Inception-ResNet-v2 model for feature vector generation ing the DFU on foot image to a remote monitor. In [20],
and utilizes the SSO algorithm for hyperparameter tuning. a novel image processing system was presented for effectual
The SSAE model is used for DFU data set classification. calculation and classifier of DFU images. Primarily, pre-
The rest of the paper is organized as follows. Section II processed was completed by cascaded fuzzy filter and then
provides the related works and section III offers the pro- non-linear partial differential equation (NPDE) based seg-
posed model. Then, section IV gives the result analysis and mentation which segmented the foot ulcer areas. Accordingly,
section V concludes the paper. the LBP was utilized for extracting the valuable features.
Afterwards, the presented hybrid GWO-CNN technique uti-
lizes these features for identifying the DFU areas. In [21],
II. LITERATURE REVIEW a unique stacked parallel convolutional layer-based network
Cassidy et al. [14] present a mobile and cloud-based structure (DFU_SPNet) was presented for performing DFU versus
for the automatic recognition of DFUs and perform an anal- normal skin classification.
VOLUME 11, 2023 127579
 S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

FIGURE 1. Workflow of SSODL-DFUDC approach.

In [22], a pre-training ResNet-50 approach and altered ture for simultaneously capturing and retaining the local and
classical-quantum approach can be employed for DFU global data of pathology images. Huang et al. [28] introduce
classification as equivalent class labels like ischaemia/non- an endwise depth domain adaptive network (DDANet) with
ischaemia and normal/abnormal. Al-Garaawi et al. [23] a combination gradient CAM and priori experience-guided
examine a CNN-based DFU classification approach where attention for improving the tumour grading solution and
it can be demonstrated that feeding a suitable feature to the interpretability by presenting the pathologist’s previous expe-
CNN technique offers a complementary efficiency to typical rience in high-magnification as depth model. Zhou et al.
RGB-based deep approaches of DFU classification tasks and [29] developed a laryngeal cancer classification network
the optimum solution was attained when either RGB images (LPCANet) that relies on a CNN and attention mechanisms.
or their texture features can be integrated and utilized as input Primary, the original histopathological images can be sequen-
to CNN. da Costa Oliveira et al. [24] present the utilisation tially cropped as patches. Huang et al. [30] introduce a novel
of DL approaches to help in the cure of DFUs. In detail, the fusion attention block network (FABNet) for addressing these
recognition of ulcers by photos taken in the patient feet. The issues.
authors present a development of Faster R-CNN utilizing data
augmentation approaches and modifications in parameter set-
tings. Venkatesan et al. [25] aimed to develop and design a III. THE PROPOSED MODEL
novel lightweight CNN approach to analyse NFU termed as In this article, we have developed a new SSODL-DFUDC
NFU-Net. technique to detect and classify DFU. The presented SSODL-
Huang et al. [26] present an endwise ViT-AMC net- DFUDC technique’s goal lies in identifying and classifying
work (ViT-AMCNet) with adaptive model fusion and DFU. Fig. 1 represents the workflow of the SSODL-
multi-objective optimizer which combines and fuses the ViT DFUDC algorithm. The figure indicates that the presented
and AMC blocks. Zhou et al. [27] examine an adaptive sparse SSODL-DFUDC technique follows three major processes:
interaction ResNet-ViT dual-branch network (ASI-DBNet). Inception-ResNet-v2 feature extraction, SSO-based hyperpa-
Initially, the authors plan the ResNet-ViT parallel infrastruc- rameter tuning, and SSAE classification.

127580 VOLUME 11, 2023

S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

A. FEATURE EXTRACTION USING OPTIMAL DL MODEL

The presented SSODL-DFUDC technique employed the
Inception-ResNet-v2 model for feature vector generation.
CNN is the algorithm of the DL algorithm. CNN is a pro-
gression of MLP developed for processing the dataset in
grid form [31]. CNN is exploited on the image dataset. The
objective of the training model is to train the ANN module
to reduce the error of predictive results of the model with the
original dataset. The convolutional layer is a convolutional
operation betwixt 2 vectors. In Eq. (1), it is convolutional of
two functions where g(x) is named the convolutional kernel
(filter) that is functioned in shifts on the vector F(x).
Z
h (x) = F (x) · g (x) = F (a) · g (x − a) (1)

FIGURE 2. Sample images a) Normal (Healthy) b) Abnormal (Ulcer).

CNN comprises two phases. Initially, group images with
feedforward. Then, use the backpropagation algorithm. Cur-
rently, the wrapping and cropping systems are performed In this work, sparrows are alienated into vigilant and dis-
to emphasize the categorised object before doing the clas- coverer followers. The location of every sparrow corresponds
sification. Next, training can be done by backpropagation to the solution. The position of each sparrow in the sparrow
and feed-forward techniques. The structure of CNN is split groups is characterized as X matrix:
into two most important parts, such as the Fully-Connected χ1,2 . . . a1,d
 

X =  ... . . . ... 
Layer and Feature Extraction Layer. Inception-ResNetV2
(2)
 
is a CNN architecture trained originally on the ImageNet
datasets, comprising 164 layers. The model has learned fea- xm,1 . . . xm,d
ture representation and is stronger for a large number of In Eq. (10), m indicates the count of sparrows, and d denotes
image types. This model accepts the input of 299 × 299 and the dimension of the variable that is enhanced. The fitness
the hybridization of two currently established networks, the function (FF) corresponds to all the sparrows are character-
modern architectural version, and residual links. There are ized as an F matrix:
sequences of filters, namely 1 × 1, 3 × 3, 5 × 5, and so 
f ([x1,1 . . . x1,d ])

on, merged with all the branch concatenation. The split-
F =  ... . 
. . . ..  (3)

transformation-mixing architecture of the initiation model
is perceived in its thick layer as a stronger representation. f ([xm,1 . . . xm,d ])
The residual connection enables the training of progressively where f ([xi,1 , · · · , xi,d ]) signifies the fitness value of i − th
DNN, leading to remarkable performance. This model was sparrows. In all the iterations, the position of discoverer will
trained and fine-tuned through TL in this work. be upgraded, as follows:

TABLE 1. Details of dataset.
 x t · exp( −i ), if R2 < ST
Xijt+1 = ij
α·N (4)
 x t + Q · L, if R ≥ ST
ij 2

where t signifies the existing amount of iterations, and N

signifies the maximum amount. α indicates a random integer,
and α ∈ (0, 1]. R2 characterizes the alarm value that is an
In this work, the SSODL-DFUDC technique utilizes the arbitrary integer, and R2 ∈ [0, 1][0, 1].ST signifies the safety
SSO algorithm for hyperparameter tuning. The SSO is a threshold and ST ∈ [O.5, 1.0].Q shows the random number
recent heuristic approach inspired by the behaviours of spar- subjected to the standard distribution, and L represents the
rows foraging and avoiding predators and characteristics of row vector whose element is equivalent to 1. The updating
group socialization [32]. The SSO benefits from strong opti- rules of the follower are shown below:
  t t

mization ability, simple architecture, easier implementation, Q · exp XW −X ι′ J
, if i > n2

few control parameters, and so on. The SSO to converge to the t+1 i2
Xij = (5)
present optimum solution is to jump towards the existing ideal x t+1 + X t − x t+1 · M ′ · L,

if i ≤ n
DB ij DB 2
solution’s locality directly. Hence the SSO technique is better
t signifies the worst location of the t−th iteration,
In Eq. (5) XW
than the particle swarm optimization, grey wolf optimization,
t+1
and gravity search algorithms with respect to robustness, and xDB signifies the location of finders with maximum
accuracy, convergence speed, and stability. fitness value at the t + 1 iteration .M signifies the 1 × d

VOLUME 11, 2023 127581

 S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

TABLE 2. Classifier outcome of SSODL-DFUDC approach with distinct measures and folds.

matrix where the element is randomly fixed to 1 or—l. M ′ = The fitness selection can be a vital element in the SSO
M T (M · M T )−1 , and M T implies the matrix was transposed. method. Solution encoding can be used to evaluate the
The position updating of vigilant is: goodness of the candidate solution. The accuracy value was
the main condition used to devise an FF.

t + γ · X t − xt ,
xGB
 if fi ̸ = fB Fitness = max (P) (7)
t+1  ijt t GB
Xi,j = X −x (6) TP
|fi −fw |+ε ,
t +k · i,j GB
if fi = fB P= (8)
xGB

TP + FP
From the expression, TP represents the true positive, and FP
t signifies the global optimum position at the t-
In Eq. (6), xGB denotes the false positive value.
th iteration, γ indicates the control step size, and γ −N (O, 1).
k shows the random value and k ∈ [−1, 1], while ε indicates B. DFU CLASSIFICATION USING THE SSAE MODEL
the constant that rises to avoid the denominator from being. In this study, the classification of DFU takes place using
fi indicates the fitness value of i − th sparrows, whereas fB the SSAE model. Hinton, in 2006, proposed an Autoen-
and fw signify the fitness values of global optimum and worst coder (AE) paradigm of ANN, and it is extensively used
sparrows, correspondingly. The superparameters that the SSO in reduction dimensionality [33]. The primary objective is
could alter include the safety value ST, the maximal amount to train the model with a similar input and output dataset.
of iterations N , and n, the number of sparrows. Furthermore, The architecture of the AE network contains a hidden layer
there are upper and lower bounds of the independent parame- (HL) with m neurons, an input layer with n neurons, and an
ter DL, several discoveries PD, and the number of vigilantes output layer with n neurons. Firstly, the n-dimension sample
SD. The independent parameter’s d dimension is determined is mapped from the input to the hidden layers that charac-
by the number of network layers and the input vector. terize an encoding method of hw1,b1 (x) ∈ Rm . Following,

127582 VOLUME 11, 2023

S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

FIGURE 3. Confusion matrices of SSODL-DFUDC approach (a-j) Fold 1-10.

FIGURE 4. TACC and VACC analysis of the SSODL-DFUDC system.

hw1,b1 (x) is decoded from the HL and mapped back to the loss function of the AE network can be determined by:
n-dimensional space Rn to recreate the input. The training
process of the AE network is to perform the backpropagation m  2
λX

1 X 1 2
of error and continuously upgrade the network parameter for JAE (W , b) = ∥x̃ − x∥22 + Wj 2
(9)
making the x̃ output to the x input, where x̃ = hw,b (x). The N 2 2
i=1 j=1

VOLUME 11, 2023 127583

 S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

FIGURE 5. TLS and VLS anlysis of SSODL-DFUDC system.

FIGURE 6. Precision-recall analysis of SSODL-DFUDC system.

where JAE (W , b) indicates the loss function of the AE net- weight matrices between jth layers and the next layer, and
work, x denotes the input of the network, x̃ shows the output λ denotes the weight of regularization. Once the amount
of the network, N represents the overall amount of input of HL nodes is lesser than the input layer, the output of
samples, ||Wj ||22 shows the regularization, Wj indicates the HL is regarded as a new feature extracted to accomplish

127584 VOLUME 11, 2023

S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

FIGURE 7. ROC analysis of the SSODL-DFUDC system.

reduction dimensionality. Once the amount of HL nodes could eventually extract the feature concealed in the input for
is larger, a sparse autoencoder (SAE) is adopted. To avoid realizing reduction dimensionality.
overfitting, SAE added sparsity limitation to the AE network
and exploited the sigmoid function as an activation function
that inhibits the neuron in the hidden layer. The suppressed IV. EXPERIMENTAL VALIDATION
state implies that the neuron output is closer to zero. For the In this section, the DFU results of the SSODL-DFUDC tech-
input x, the activation degree of jth neurons in the HL is hj (x), nique are tested using a dataset from the Kaggle repository
then the average activation degree of these neurons in the [34]. The dataset includes 844 samples, with 410 abnormal
entire training samples S = {x(i) }N is ρ̂ 1 PN and 434 standard samples, as demonstrated in Table 1. Fig. 2
i=1 j = N i=1 j (x(i) ).
h
To guarantee that the neuron is suppressed, we must make showcases the sample images of normal and abnormal.
the average activation degree closer to the sparse variable ρ, Table 2 represents the overall DFU detection results of
which is generally closer to zero.A penalty factor is added to the SSODL-DFUDC technique under different folds. The
the loss function of AE to accomplish the sparsity. The loss results showed that the SSODL-DFUDC technique has
function of the SAE network is formulated by: enhanced results under all folds. For instance, with fold-
1, the SSODL-DFUDC technique reaches average accubal ,
J
X precn , and recal of 99.17%, 99.17%, and 99.17%, respec-
Jsparse (W , b) = JAE (W , b) + β KL ρ∥ρ̂j


(10)
tively. Meanwhile, with fold-5, the SSODL-DFUDC system
j=1
gains average accubal , precn , and recal of 99.17%, 99.17%,
In Eq. (10), ρ̂j indicates the average activation degree of and 99.17%, correspondingly. Eventually, with fold-10, the
each neuron with every training sample, ρ denotes the sparse SSODL-DFUDC approach attains average accubal , precn ,
variable to be set. KL(ρ||ρ̂j ) represents the relative entropy and recal of 99.55%, 97.51%, and 97.55%, correspondingly.
between ρ and ρ̂j , viz., the KL divergence that value increases The TACC and VACC of the SSODL-DFUDC approach
monotonically as the difference between ρ and ρ̂j .β denotes are investigated on DFU performance in Fig. 4. The figure
the weight of KL divergence. Once ρ and ρ̂j are equivalent, pointed out that the SSODL-DFUDC system has shown
then KL(ρ||ρ̂j ) is equivalent to for getting its minimal value. improved performance with improved values of TACC and
Since ρ is generally closer to zero, ρ̂ of each neuron in the HL VACC. It is perceptible that the SSODL-DFUDC system has
would be controlled to be within a smaller range afterwards reached maximal TACC outcomes.
training, and the sparse representation of the input signal is The classification results of the SSODL-DFUDC tech-
lastly attained. The SSAE is encompassed by a single-layer nique are inspected in the form of a confusion matrix in Fig. 3.
SAE network, with the output of SAE of the preceding layer, The results implied that the SSODL-DFUDC technique has
which acts as an input of SAE of the following adjacent layer. successfully recognized the presence of normal and abnormal
By training the SAE network layer-wise, the SSAE network samples.

VOLUME 11, 2023 127585

 S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

TABLE 3. Comparative analysis of SSODL-DFUDC system with existing algorithms [35].

The TLS and VLS of the SSODL-DFUDC method are the improved efficiency of the SSODL-DFUDC approach
tested on DFU performance in Fig. 5. The figure states compared to existing DL techniques. In the future, the
that the SSODL-DFUDC technique has exposed improved SSODL-DFUDC algorithm’s performance can be improved
performance with the lowest values of TLS and VLS. It is by advanced DL classification models. In addition, weakly
observable that the SSODL-DFUDC model has resulted in supervised learning techniques can be explored to reduce
reduced VLS outcomes. the dependency on precise pixel-level annotations. Instead of
An evident precision-recall study of the SSODL-DFUDC relying solely on expert-labeled ulcer regions, weakly super-
methodology in the test database is described in Fig. 6. The vised learning can utilize weak labels, such as image-level or
figure implied that the SSODL-DFUDC system has led to bounding box annotations, making it more scalable and easier
superior values of precision-recall values in two classes. to collect data.
A detailed ROC study of the SSODL-DFUDC algorithm
in the test database is exposed in Fig. 7. The outcome stated ACKNOWLEDGEMENT
the SSODL-DFUDC system has displayed its capability to Princess Nourah bint Abdulrahman University Researchers
classify two class labels. Supporting Project number (PNURSP2023R300), Princess
To illustrate the enhanced performance of the SSODL- Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.
DFUDC technique, a widespread comparison study is made
in Table 3 and Fig. 8 [35]. The experimental values
REFERENCES
indicated that the AlexNet model reaches poor perfor-
[1] S. Wang, J. Wang, M. X. Zhu, and Q. Tan, ‘‘Machine learning for the
mance. At the same time, the DFUNet-KNN, DFUNet, and prediction of minor amputation in university of Texas grade 3 diabetic foot
DFUNet-SVM models have attained slightly enhanced out- ulcers,’’ PLoS ONE, vol. 17, no. 12, Dec. 2022, Art. no. e0278445.
comes. Along with that, the VGG16 and GoogleNet models [2] M. Swerdlow, L. Shin, K. D’Huyvetter, W. J. Mack, and D. G. Armstrong,
have obtained moderately improved performance. Although ‘‘Initial clinical experience with a simple, home system for early detection
and monitoring of diabetic foot ulcers: The foot selfie,’’ J. Diabetes Sci.
the EfficientNet model results in reasonable performance, Technol., vol. 17, no. 1, pp. 79–88, Jan. 2023.
the SSODL-DFUDC technique outperforms the existing ones [3] M. Goyal, N. D. Reeves, S. Rajbhandari, N. Ahmad, C. Wang, and
with an increased accuy of 99.29%, precn of 99.29%, recal M. H. Yap, ‘‘Recognition of ischaemia and infection in diabetic foot
ulcers: Dataset and techniques,’’ Comput. Biol. Med., vol. 117, Feb. 2020,
of 99.29%, and F1score of 99.29%. These results assured the Art. no. 103616.
improved performance of the SSODL-DFUDC technique in [4] J. Yogapriya, V. Chandran, M. G. Sumithra, B. Elakkiya, A. S. Ebenezer,
the classification process. and C. S. G. Dhas, ‘‘Automated detection of infection in diabetic foot ulcer
images using convolutional neural network,’’ J. Healthcare Eng., vol. 2022,
pp. 1–12, Apr. 2022.
V. CONCLUSION [5] K. S. Chan and Z. J. Lo, ‘‘Wound assessment, imaging and monitoring
In this article, we have developed a novel SSODL-DFUDC systems in diabetic foot ulcers: A systematic review,’’ Int. Wound J., vol. 17,
no. 6, pp. 1909–1923, Dec. 2020.
system to detect and classify DFU. The presented SSODL-
[6] B. Cassidy, C. Kendrick, N. D. Reeves, J. M. Pappachan, C. O’Shea,
DFUDC technique’s goal lies in identifying and classifying D. G. Armstrong, and M. H. Yap, ‘‘Diabetic foot ulcer grand challenge
DFU. To accomplish this, the presented SSODL-DFUDC 2021: Evaluation and summary,’’ in Diabetic Foot Ulcers Grand Chal-
technique has applied the Inception-ResNet-v2 model for lenge. Cham, Switzerland: Springer, Sep. 2021, pp. 90–105.
[7] H. Pan, H. Peng, Y. Xing, L. Jiayang, X. Hualiang, T. Sukun, and F. Peng,
feature vector generation. In addition, the SSODL-DFUDC ‘‘Breast tumour grading network based on adaptive fusion and microscopic
technique utilizes the SSO algorithm for hyperparameter imaging,’’ Opto-Electron. Eng., vol. 50, no. 1, 2023, Art. no. 220158.
tuning purposes. Moreover, the classification of DFU takes [8] Y. Wang, F. Luo, X. Yang, Q. Wang, Y. Sun, S. Tian, P. Feng, P. Huang,
place using the SSAE model. The experimental results of the and H. Xiao, ‘‘The Swin-transformer network based on focal loss is used
to identify images of pathological subtypes of lung adenocarcinoma with
SSODL-DFUDC technique take place using the DFU dataset. high similarity and class imbalance,’’ J. Cancer Res. Clin. Oncol., vol. 149,
The comprehensive experimental outcomes demonstrate no. 11, pp. 8581–8592, Sep. 2023.
127586 VOLUME 11, 2023
S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

[9] P. Huang, X. Tan, C. Chen, X. Lv, and Y. Li, ‘‘AF-SENet: Classification [28] X. Zhou, C. Tang, P. Huang, S. Tian, F. Mercaldo, and A. Santone,
of cancer in cervical tissue pathological images based on fusing deep ‘‘ASI-DBNet: An adaptive sparse interactive ResNet-vision transformer
convolution features,’’ Sensors, vol. 21, no. 1, p. 122, 2020. dual-branch network for the grading of brain cancer histopathological
[10] C. Du, Y. Li, P. Xie, X. Zhang, B. Deng, G. Wang, Y. Hu, M. Wang, images,’’ Interdiscipl. Sci., Comput. Life Sci., vol. 15, no. 1, pp. 15–31,
W. Deng, D. G. Armstrong, Y. Ma, and W. Deng, ‘‘The amputation and 2023, doi: 10.1007/s12539-022-00532-0.
mortality of inpatients with diabetic foot ulceration in the COVID-19 [29] P. Huang, X. Zhou, P. He, P. Feng, S. Tian, Y. Sun, F. Mercaldo,
pandemic and postpandemic era: A machine learning study,’’ Int. Wound A. Santone, J. Qin, and H. Xiao, ‘‘Interpretable laryngeal tumor grading
J., vol. 19, no. 6, pp. 1289–1297, Oct. 2022. of histopathological images via depth domain adaptive network with inte-
[11] O. Güley, S. Pati, and S. Bakas, ‘‘Classification of infection and ischemia gration gradient CAM and priori experience-guided attention,’’ Comput.
in diabetic foot ulcers using VGG architectures,’’ in Diabetic Foot Ulcers Biol. Med., vol. 154, Mar. 2023, Art. no. 106447.
Grand Challenge. Cham, Switzerland: Springer, Sep. 2021, pp. 76–89. [30] X. Zhou, C. Tang, P. Huang, F. Mercaldo, A. Santone, and Y. Shao,
[12] J. Tulloch, R. Zamani, and M. Akrami, ‘‘Machine learning in the pre- ‘‘LPCANet: Classification of laryngeal cancer histopathological images
vention, diagnosis and management of diabetic foot ulcers: A systematic using a CNN with position attention and channel attention mechanisms,’’
review,’’ IEEE Access, vol. 8, pp. 198977–199000, 2020. Interdiscipl. Sci., Comput. Life Sci., vol. 13, no. 4, pp. 666–682, Dec. 2021.
[13] W. Ren, A. H. Bashkandi, J. A. Jahanshahi, A. Q. M. AlHamad, [31] P. Huang, X. Tan, X. Zhou, S. Liu, F. Mercaldo, and A. Santone, ‘‘FABNet:
D. Javaheri, and M. Mohammadi, ‘‘Brain tumor diagnosis using a step- Fusion attention block and transfer learning for laryngeal cancer tumor
by-step methodology based on courtship learning-based water strider grading in P63 IHC histopathology images,’’ IEEE J. Biomed. Health
algorithm,’’ Biomed. Signal Process. Control, vol. 83, May 2023, Informat., vol. 26, no. 4, pp. 1696–1707, Apr. 2022.
Art. no. 104614. [32] Y. Guo, D. Yang, Y. Zhang, L. Wang, and K. Wang, ‘‘Online estimation
[14] B. Cassidy, N. D. Reeves, J. M. Pappachan, N. Ahmad, S. Haycocks, of SOH for lithium-ion battery based on SSA-elman neural network,’’
D. Gillespie, and M. H. Yap, ‘‘A cloud-based deep learning framework for Protection Control Mod. Power Syst., vol. 7, no. 1, pp. 1–17, Dec. 2022.
remote detection of diabetic foot ulcers,’’ IEEE Pervasive Comput., vol. 21, [33] B. Li, S. Liang, D. Chen, and X. Li, ‘‘A decision-making method for
no. 2, pp. 78–86, Apr. 2022. air combat maneuver based on hybrid deep learning network,’’ Chin. J.
[15] K. Munadi, K. Saddami, M. Oktiana, R. Roslidar, K. Muchtar, M. Melinda, Electron., vol. 31, no. 1, pp. 107–115, Jan. 2022.
R. Muharar, M. Syukri, T. F. Abidin, and F. Arnia, ‘‘A deep learning method [34] Diabetic Foot Ulcer (DFU). Accessed: Jul. 14, 2023. [Online]. Available:
for early detection of diabetic foot using decision fusion and thermal https://www.kaggle.com/datasets/laithjj/diabetic-foot-ulcer-dfu
images,’’ Appl. Sci., vol. 12, no. 15, p. 7524, Jul. 2022. [35] P. N. Thotad, G. R. Bharamagoudar, and B. S. Anami, ‘‘Diabetic foot ulcer
[16] M. H. Alshayeji, S. C. Sindhu, and S. Abed, ‘‘Early detection of diabetic detection using deep learning approaches,’’ Sensors Int., vol. 4, Jan. 2023,
foot ulcers from thermal images using the bag of features technique,’’ Art. no. 100210.
Biomed. Signal Process. Control, vol. 79, Jan. 2023, Art. no. 104143.
[17] G. D’Angelo, D. Della-Morte, D. Pastore, G. Donadel, A. De Ste-
fano, and F. Palmieri, ‘‘Identifying patterns in multiple biomarkers
to diagnose diabetic foot using an explainable genetic programming-
based approach,’’ Future Gener. Comput. Syst., vol. 140, pp. 138–150,
Mar. 2023. SUNNAM NAGARAJU received the B.Tech.
[18] L. Alzubaidi, M. A. Fadhel, S. R. Oleiwi, O. Al-Shamma, and J. Zhang, degree from JNTUH, Hyderabad, in 2007, and
‘‘DFU_QUTNet: Diabetic foot ulcer classification using novel deep con- the M.Tech. degree from NIT, Trichy, Tamil
volutional neural network,’’ Multimedia Tools Appl., vol. 79, nos. 21–22, Nadu. He was an Assistant Professor with Lovely
pp. 15655–15677, Jun. 2020. Professional University, Jalandhar, Punjab, India,
[19] M. Goyal, ‘‘Novel computerised techniques for recognition and anal- from 2010 to 2016. He was also a Plant Engineer
ysis of diabetic foot ulcers,’’ Ph.D. thesis, Manchester Metropoli-
with OCTL, Hyderabad. He is currently an Asso-
tan Univ., 2019. [Online]. Available: https://e-space.mmu.ac.uk/625105/
ciate Professor with MLRIT, Hyderabad, in 2010.
1/Thesis_Manu_Revised.pdf
[20] T. Arumuga Maria Devi and R. Hepzibai, ‘‘Clinical assessment of dia-
betic foot ulcers using GWO-CNN based hyperspectral image processing
approach,’’ IETE J. Res., 2022, doi: 10.1080/03772063.2022.2099469.
[21] S. K. Das, P. Roy, and A. K. Mishra, ‘‘DFU_SPNet: A stacked parallel con-
volution layers based CNN to improve diabetic foot ulcer classification,’’
ICT Exp., vol. 8, no. 2, pp. 271–275, Jun. 2022.
[22] J. Amin, M. A. Anjum, A. Sharif, and M. I. Sharif, ‘‘A modified classical- KOLLATI VIJAYA KUMAR received the Ph.D.
quantum model for diabetic foot ulcer classification,’’ Intell. Decis. degree in computer science and engineering
Technol., vol. 16, no. 1, pp. 23–28, Apr. 2022. from Karpagam University, Coimbatore. He has
[23] N. Al-Garaawi, R. Ebsim, A. F. H. Alharan, and M. H. Yap, ‘‘Diabetic 15 years of teaching experience. He is currently
foot ulcer classification using mapped binary patterns and convolu- an Associate Professor with the Department of
tional neural networks,’’ Comput. Biol. Med., vol. 140, Jan. 2022, CSE, GITAM School of Technology, GITAM Uni-
Art. no. 105055. versity, Visakhapatnam, Andhra Pradesh, India.
[24] A. Oliveira, A. B. de Carvalho, and D. Dantas, ‘‘Faster R-CNN approach He has published 24 articles in international
for diabetic foot ulcer detection,’’ in Proc. 16th Int. Joint Conf. Comput. journals. His research interests include wireless
Vis., Imag. Comput. Graph. Theory Appl., 2021, pp. 677–684. networking, big data, data analytics, network secu-
[25] C. Venkatesan, M. G. Sumithra, and M. Murugappan, ‘‘NFU-Net: An auto- rity, cloud computing, and information security.
mated framework for the detection of neurotrophic foot ulcer using deep
convolutional neural network,’’ Neural Process. Lett., vol. 54, no. 5,
pp. 3705–3726, Oct. 2022.
[26] C. A. Ferreira, T. Melo, P. Sousa, M. I. Meyer, E. Shakibapour, P. Costa, and
A. Campilho, ‘‘Classification of breast cancer histology images through
transfer learning using a pre-trained inception resnet v2,’’ in Proc. Int.
Conf. Image Anal. Recognit. Cham, Switzerland: Springer, May 2018, B. PRAMEELA RANI received the B.Tech. degree from the Aditya College
pp. 763–770. of Engineering and Technology (ACET), Suramplaem, in 2014, and the
[27] P. Huang, P. He, S. Tian, M. Ma, P. Feng, H. Xiao, F. Mercaldo, M.Tech. degree from the Aditya College of Engineering (Autonomous),
A. Santone, and J. Qin, ‘‘A ViT-AMC network with adaptive model fusion Surampalem. She is currently an Assistant Professor with the Aditya
and multiobjective optimization for interpretable laryngeal tumor grading College of Engineering. She was an Instructor in engineering colleges,
from histopathological images,’’ IEEE Trans. Med. Imag., vol. 42, no. 1, from 2014 to 2021, on various technologies. Her current research interests
pp. 15–28, Jan. 2023. include machine learning models and deep learning mechanisms.

VOLUME 11, 2023 127587

 S. Nagaraju et al.: Automated Diabetic Foot Ulcer Detection and Classification Using Deep Learning

E. LAXMI LYDIA is currently a Professor in com- IMEN FILALI received the master’s degree in networks and distributed
puter science engineering with GMRIT. She is also systems and the Ph.D. degree from the University of Nice Sophia-Antipolis,
a Big Data Analytics Online Trainer with the Inter- France, in 2007 and 2011, respectively. She has been an Assistant Pro-
national Training Organization. She has presented fessor with the National Engineering School of Gabes, Tunisia, since
various webinars on big data analytics. She is also 2011. From December 2014 to September 2015, she was a Postdoctoral
with the Government DST Funded Project. She is Researcher with the Swiss Federal Institute of Technology of Lausanne
also certified by the Microsoft Certified Solution (EPFL) working within the Distributed Information Systems Laboratory
Developer (MCSD). She is the author of the Big (LSIR). Since January 2018, she has been with the Department of Computer
Data Analytics book. She holds a patent. She has Science, College of Computer and Information Sciences, Princess Nourah
published ten research papers in international con- Bint Abdulrahman University, as an Assistant Professor. Her main research
ference proceedings. She has published more than 100 research articles in interests include cloud computing, big data analytics, artificial intelligence,
international journals in Big Data Analytics and Data Science Journal. and computer vision.

FATEN KHALID KARIM received the Ph.D. degree in computing and infor-
mation technology from Flinders University, SA, Australia. She is currently
an Assistant Professor with the Department of Computer Sciences, College
of Computer and Information Sciences, Princess Nourah Bint Abdulrah-
man University, Riyadh, Saudi Arabia. Her research interests include cloud
MOHAMAD KHAIRI ISHAK received the B.Eng. computing and information technology. She has published several research
degree in electrical and electronics engineering articles in her field.
from International Islamic University Malaysia
(IIUM), Malaysia, the M.Sc. degree in embed-
ded system from the University of Essex, U.K.,
and the Ph.D. degree from the University of Bris-
tol, U.K. He is currently a Registered Graduate SAMIH M. MOSTAFA received the bachelor’s
Engineer with the Board of Engineers Malaysia and M.Sc. degrees in computer science from
(BEM). He is also a Senior Lecturer in mechatron- the Computer Science-Mathematics Department,
ics engineering with the School of Electrical and Faculty of Science, South Valley University, in
Electronic Engineering, Universiti Sains Malaysia (USM). His background 2004 and 2010, respectively, and the Ph.D. degree
includes outstanding teaching experience with Universiti Sains Malaysia, in computer science from the Advanced Infor-
instructing students to stimulate engineering information interest and reten- mation Technology Department, Graduate School
tion while invigorating classes through the use of new technologies and of Information Technology, Kyushu University,
models. His research interests include embedded systems, real-time control Japan, in 2017. His research interests include
communications, and Internet of Things (IoT). Emphasis is given toward the machine learning and CPU scheduling. He is cur-
development of theoretical and practical methods which can be practically rently a fellow of the Academy of Scientific Research and Technology
validated. Recently, significant research effort has been directed toward (ASRT), Egypt.
important industrial issues of embedded networked control systems and IoT.

127588 VOLUME 11, 2023