Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)

IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

Detection and Classification of Skin Cancer using

Binary Classifier, Residual Network, and
Convolutional Neural Network
Urmila Pilania1 Manoj Kumar2
Computer Science & Technology Computer Science & Technology
Manav Rachna University Manav Rachna University
Faridabad, India Faridabad, India
[email protected] [email protected]
2024 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS) | 979-8-3503-7999-0/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICSCSS60660.2024.10625032

Priyam Garg3 Reet Kaur4

Computer Science & Technology Computer Science & Technology
Manav Rachna University Manav Rachna University
Faridabad, India Faridabad, India
[email protected] [email protected]

Abstract: Skin cancer is a very dangerous form of cancer as results in the detection of any type of disease in images.
it transfers from one part of the body to the other part. It CNN has made a remarkable achievement in feature
can occur in any age group but is very common in old retrieval and pattern recognition [2]. ResNet design is
people. For effective treatment and the best possible very good at addressing the challenges that arise during
outcome, early detection of skin cancer is necessary.
Traditional techniques are sometimes not able to detect
the training phase of deep networks. ResNet is very useful
properly due to the lack of expertise. These methods also because of its deep layered architecture and skip
cause discomfort, and risk of compilation, and require more connections for learning features of medical images.
recovery time compared to the traditional methods. In the Binary classifier helps in differentiating between cancers.
work, the authors proposed a hybrid model for skin cancer
detection and classification using a Binary Classifier, Few of the features of all three techniques are [4]:
Residual Network, and Convolutional Neural Network.
The research work first uses a binary classifier to distinguish CNN is the structure that is very helpful in detection,
between malignant and benign cancer. Then, Residual classification, segmentation, and image generation. By
Network is used to understand its deep architecture and skip
connections, for retrieving features from input images.
using filters CNN extracts features from the input images.
Finally, the Convolution Neural Network (CNN) is used Pooling layers retain the most important information and
specifically to detect skin cancer, leveraging its architectural delete the unwanted data by reducing the spatial
design and training strategies. The publicly available Kaggle dimensions. The activation function adds non-linearity to
dataset is utilized for the research work. For assessment of the CNN which enables learning the relationship of data.
the work accuracy and time have been calculated. The CNN represents the features in hierarchical form the
experimental results authenticate the effectiveness of the lower layers have simple information like edges and
model in precisely detecting and classifying skin cancer. textures. The higher layers have information relevant to
the task the user wants to perform. In CNN layers are
Keywords: Skin Cancer, Residual Network (ResNet), connected so it aggregates the features generated by the
Convolutional Neural Network (CNN), Binary Classifier,
Detection, Classification, Images.
previous layer and produces output. CNN has features of
data augmentation which increase the diversity in given
I. INTRODUCTION dataset. ResNet provides very deep training of the residual
blocks. It provides ideal mapping by skipping the blocks
Skin cancer is a global health issue nowadays, with its
as and when required. ResNet uses the concept of global
occurrence over the years in all age people. Traditional
average polling so it decreases the constraints resulting in
methods like visual examination and dermoscopy has
optimal outcome. Binary classifiers work on decision
been utilized in detecting cancer. The outcome of the
boundaries that segregate the different classes. The binary
traditional methods depends on the proficiency of the
classifier uses supervised learning where learning takes
clinician. Additionally, the number of skin cancer are
place through labelled data.
increasing day by day because of that timely diagnosis is
required and we need some automated tool to detect the
disease. Deep Learning (DL) methods have encouraging

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1115

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

Fig. 1. Region-wide Distribution

Figure 1 characterizes the region-wide distribution of The back part of the human body is highly affected and
cancer. The X-axis signifies the region and the y-axis the acral is the least affected area as per Figure 1. The X-
signifies patients affected from the cancer on different axis in Figure 2 represents age of the cancer patient and
parts of the body. the Y-axis signifies the count of patients affected by the
disease. The age group 40 to 50 is the highly pretentious
group with the cancer disease. Figure 3 illustrates the
organization of cancer into major seven categories. The
most occurring cancer is melanocytic nevi and the least
occurring skin cancer is dermatofibroma. Figure 4
represents the gender-wise data for skin cancer
identification and classification. The X-axis signifies the
gender-wise data and the Y-axis signifies count of
patients. Males are more affected by skin cancer matched
to the female.

Fig. 2. Age-wise Distribution of Skin Cancer

Fig. 3. Classification of Cancer

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1116

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

dataset and ISIC 2018 dataset, accuracy of 98.35% was

attained [1].
Melanoma is a life-threatening tumor and currently it has
been on the rise. If it is treated timely then the life of the
patient could be saved. The traditional methods of
Melanoma disease identification are slow and may have
the wrong outcome. So the automated melanoma
identification technique is required for timely treatment.
In the work, neural network method was proposed for the
identification and classification of melanoma disease.
Dermoscopic images were pre-processed using the
Maximum Gradient Intensity methods for the removal of
hairs from the images. Then Otsu thresholding technique
was used to segregate skin lesions and for segmentation.
The LPB was applied to train the NN model. Using ISIC
and PH2 dataset accuracy of 97.7% was attained by the
proposed work [3].
It is very hard to detect melanoma with traditional
techniques so the same advanced automated melanoma
Fig. 4. Gender wise Skin Cancer detection techniques are required so that timely treatment
The research contribution for research work is: can be started for the patient. Authors in paper, proposed a
• In research work, a hybrid method is designed to ML method for the identification and recognition of
detect and classify skin cancer. melanoma. Computational techniques were required to
• The proposed method carries high accuracy and compute, analyze, and process the input image. The
reliability. images were pre-processed to remove the unwanted data
• The publically accessible dataset was utilized for from the image. The required features were retrieved
identification and classification skin cancer. using CNN and recognition was done using softmax
classifier. The simulation of the proposed work proved
• The performance of work is measured using
that it is a more efficient and reliable technique equated to
assessment parameters of accuracy, precision,
the existing techniques [5].
and time taken by the model.
The automated disease identification technique reduces
• The literature survey helps in finding the
the time and workload of the user. These techniques are
limitations of the existing work which helps in
also having good performance results like accuracy and
finding future direction for the researchers and
other performance metrics. Artificial intelligence plays a
academicians.
significant role in medicine these days. The growth of
The paper is organized into 5 sections. Section 1 is about
smart electronic gadgets in this modern era for DL and
the introduction of skin cancer and the techniques that we
machine learning is very helpful for professionals. In the
proposed for the identification and classification. Section
proposed work neural networks, CNN, Support Vector
2 provides summary of related research for detection and
Machine (SVM), deep learning techniques, and fuzzy
classification of skin cancer. Section 3 defines
logic techniques were reviewed to design the automated
methodology adopted for the work, like dataset used,
melanoma identification model [6]. These days melanoma
CNN model, binary classifier, ResNet, and evaluation
disease become critical issue worldwide. The
metrics. Section 4 represents the simulation of work with
dermoscopic images could be utilized to recognize the
results and discussions. Section 5 concludes the research
disease using artificial intelligence techniques. In the
work and its scope in future.
research work CNN was utilized by the authors to identify
II. LITERATURE REVIEW and classify the disease. CNN is very good in feature
extraction and leading to improved classification. The
These days’ DL models have become very advanced in proposed work also used an Optimized Stacked Sparse
medical imaging and analysis. The identification and Auto Encoder (OSSAE) based feature extractor with
recognition of skin cancer at primary stage is important to Backpropagation Neural Network (BPNN), named the
save the life of the patient. For the research work, DL and OSSAE-BPNN technique. The publically accessible
Artificial Intelligence (AI) algorithms was used by dataset was utilized in the work for the identification and
authors to detect melanoma. Two datasets PH2 and ISIC classification of the disease. For assessment of work
2018 were used having dermoscopy images. The Local performance metrics showed good results [8].
Binary Pattern (LBP) and Gray Level Co-occurrence DL technique has attained important progress in
Matrix (GLCM) algorithms were applied to the retrieval melanoma diagnosis. These methods are properly utilized
of features. An Artificial Neural Network (ANN) was for clinical purposes. It has been witnessed by the
applied to these retrieved features to process the literature that the accuracy of the DL methods is reaching
information. The simulation of the work proved the 90 % but still some cases left unidentified. ISIC 2019
efficiency and reliability. The assessment metrics were dataset having 487 dermoscopic images has been utilized
calculated to measure the quality of work. The ANN for the identification and recognition of disease. Authors
technique attained a good accuracy of 97.50% for the PH2 have trained multiple transfers learned segmentation for

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1117

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

the identification of irregular networks in the training set. • In limited access areas, it is challenging for
For classification purposes dataset of 500 melanomas and dermatologists to properly diagnose skin cancer.
500 benign images was also utilized during training. The • The class inequity in the dataset might cause
accuracy and recall of melanoma were found to be high inaccurate in skin cancer identification and
related to the benign [9]. recognition.
The reduction of the ozone layer is causing skin-related • The automated techniques developed by different
issues for humans along with the environment issues. researchers are not general for a diverse
Exposure to ultraviolet rays harms human skin resulting in population.
skin disease. The timely identification results in timely
treatment of the disease leading to the good outcome. The III. PROPOSED METHOD
AI and DL models were combined for better results. The This research work focuses on leverage the proficiencies
publically available dataset having 8598 images has been of DL models. Binary classifier is integrated with ResNet
used by the authors for the identification and classification and CNN, to improve the detection and classification of
of the melanoma disease. AlexNet, MobileNet, ResNet, skin cancer from dermoscopic images. The Kaggle dataset
VGG16, and VGG19, have been studied and it was found has been used for the dermoscopic images. The dataset is
that as MobileNet model achieved best then other publically available having 10015 images and all the
mentioned techniques. The accuracy attained for the images are labelled properly. An Epoch of 700 sizes at a
work, using MobileNet is 84.94% [10]. learning rate of 0.05 has been selected. Batches of size 32
After the literature survey a lot of issues have been listed are created for the work. Batch size depends on many
as follows [11][12]: parameters such as computational resources; size of
• Skin cancer resembles benign moles or lesions so dataset, and architecture of research work, etc. Batch size
for professionals accurate identification of cancer varies with the memory size of the GPU used. The larger
is very challenging. the batch size requires more amount of memory. Large
• The results of the different professionals vary for batch size reduces the transfer of memory from CPU to
the same dataset and the technique used. GPU. The batch size in general could be 16 to 128
• Skin cancer can vary in color, size, shape, and samples. Figure 5 represents the working of the DL
texture. Its location could also vary resulting in technique in general.
issues for the professionals.
• Skin cancer appearance can vary over time,
leading to missed diagnoses during early
screenings.

Fig. 5. General Cancer Detection Flow [7]

A publicly accessible dataset from Kaggle is chosen for pre-processing all the images are scaled in the same range
the work. The dataset has 10015 images; these images are so that during training they can contribute equally. New
labeled with age, sex, image ID, lesion ID, and type. The features might be added to the images or transformation
collected data is then pre-processed by adding missing might be used to the images. The dataset is split: training,
values, and outliers, and removing the unwanted data. In testing, and validation sets. CNN model is selected for

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1118

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

further processing like to train the pre-processed data. The Deployment and Clinical Validation: After the
trained data is fed to the system for prediction. The loss verification and validation of research the technique is
function calculates the difference between predicted and deployed for the real- world applications [17].
actual values. Model performance is accessed on the The proposed research work will provide significant
validation set for tuning hyper parameters and prevents advancement in the premature recognition and diagnosis
the overfitting of data. Testing of the technique is done on of skin cancer, eventually contributing to enhanced patient
different sets to evaluate the generalized model results and reducing the load of deadly diseases.
performance. Finally, the technique is installed for
production to predict new datasets. Finally, for IV. RESULTS AND DISCUSSION
maintenance purposes continuous monitoring is required, The section presents outcomes produced by the hybrid
and for adaption of the changes needs to be done from model. We have combined CNN, ResNet, and binary
time to time. classifiers to detect and classify skin cancer. The
The proposed methodology includes many key steps experiment has been attained on the dermoscopic images
outlined below: dataset from the Kaggle. The dataset is decomposed:
Data Acquisition and Pre-processing: A Kaggle dataset training, validation, and testing sets into the ratio of
having 10015 dermoscopic images has been used. It is a 70:15:15. ResNet is utilized for feature retrieval, and the
publically accessible dataset. Then pre-processing of binary classifier is used for supervisory. The discussion of
dataset has been done to change image size, and outcomes for the detection and recognition of cancer
information perceptibility has been enhanced [13]. utilizing a hybrid model could normally include an
Model Architecture Design: Pre-trained ResNet exploration of the assessment metrics of the designed
architecture used to retrieve the features from automatic model, possible demerits, and opportunities for
dermoscopic images. The network performed well in the future research.
extraction of features from medical images. It also helps The results of the designed automatic model could be
in classification of the skin cancer. The binary classifier accessible in form of several assessment metrics such
consists of fully connected layers monitored by a sigmoid accuracy, recall, specificity, time, precision, and F1-score.
activation function to output a probability score The mentioned metrics deliver perceptions into the
representing the likelihood of skin cancer existence [14]. technique's proficiency to accurately classify skin cancer
Training Strategy: The method is trained using an SL into actinic keratoses, benign keratosis, melanocytic nevi,
method on the pre-processed dataset. In the work, transfer dermatofibroma, basal cell carcinoma, and vascular
learning has been used, initializing the ResNet model with lesions. If the research work achieves high accuracy in
pre-trained weights in the dataset. Fine-tuning would help considerable time then its effectiveness could be verified.
in acclimating the model to the particular task of skin The high accuracy in considerable time showed that it
cancer recognition [15]. could be used as a generalized model for clinical
Model Evaluation: The trained model is tested and applications.
validated on different datasets to measure its performance.
For measuring the performance accuracy, recall, and time
has been calculated [16].

Fig. 6. Detection and Classification of Skin Disease

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1119

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

Fig. 7. Detection and Classification of Skin Cancer

Figures 6 and 7 represent the detected and classified skin varying from year 0 to 80. Figure 8 represents the skin
cancer. In the dataset, 7 types of skin cancer have been cancer distribution, the cancer types 0, 1, 3, and 5 which
detected and classified. The most occurring skin cancer is are Melanocytic nevi, dermatofibroma, basal cell
melanocytic nevi and the least occurring skin cancer is carcinoma, and vascular lesions are not very prevalent
dermatofibroma. The dataset comprises patient details below the age of 20 years.

Fig. 8. Skin Cancer Distribution

type of situation, other assessment metrics like recall,

Accuracy is the part of accurately classified instances precision, and F1-score could be used [18].
divided by the overall instances. It is the maximum Accuracy
utilized performance metric used for assessment of Number of Accurate Prediction
machine learning techniques. But in some cases like = … eq. 1
Total number of Predictions
imbalanced classes, it is not able to work properly. In this

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1120

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

Fig. 9. Accuracy and Loss Graph

resulting in timely treatment of patients. The timely

Figure 9 represents the accuracy of training and validation preliminary screening significantly decreases the
for the proposed model along with loss for training and computational burden and confirms that only relevant
validation. At the time of training of the data existing in cases continue to more severe analysis. The binary
the dataset, the model acquires from training data by classifier efficiently filtered the malignant and benign
adjusting weights and bias for minimization of loss cancer utilizing the retrieved features. The Residual
function. The training loss calculates the error present in Network allows for learning the complex patterns and
the predicted value and the target value. The lower the features from the pre-processed images. CNN permits the
value of training loss higher the value of accuracy in the retrieval of the hierarchical representation of the input
model. Validation loss calculates the error present in the image. The integration of these techniques improves the
models predicted and the real target values on a distinct accuracy of skin cancer detection and classification. The
validation dataset that the model has not perceived during accuracy of skin cancer detection and classification of the
the training phase. The validation set measure the proposed model is calculated at 80%. The time taken to
generality of technique or we can say how it handles the detect and classify the skin cancer is also reasonable
new data. A reduced training loss in integration with a which is 43 minutes.
reduced validation loss specifies that the designed model In conclusion, the integrated technique is robust and
is learning valuable patterns from the data without efficient for skin cancer detection and classification with
overfitting. The accuracy for research work is found to be the probability to significantly add healthcare
approximately 80% at 100 epochs. The total time taken to professionals to early diagnosis and treatment planning.
detect and classify the skin cancer is 43 minutes. The Continued growth in the medical field has the potential to
technique is interoperable on any platform. It could be further improve the accuracy of skin cancer detection and
utilized on any dataset of skin cancer disease. classification, resulting in enhanced patient outcomes and
Along with its utilization, every automated techniques saving lives.
also has some challenges like the requirement of a large
size dataset, the biases in data, and other issues in
generalizing the automated model to diverse skin types, REFERENCES
and uncertainties in the interpretability of the model's
conclusions. The mentioned issues also help in framing [1]. Alsaade, F. W., Aldhyani, T. H., & Al-Adhaileh, M. H. (2021).
the future directions for the proposed model. Overall, a Developing a recognition system for diagnosing melanoma skin
thorough discussion of results for the detection and lesions using artificial intelligence algorithms. Computational and
mathematical methods in medicine, 2021, 1-20.
classification of skin cancer using research work which [2]. Kumar, S., Pilania, U., & Nandal, N. (2023). A systematic study of
delivers insights into the prospective of these techniques artificial intelligence-based methods for detecting brain
to advance healthcare outcomes, while also highlighting tumors. Informatics and Automation, 22(3), 541-575.
the issues and opportunities for further improvements in [3]. Tumpa, P. P., & Kabir, M. A. (2021). An artificial neural network-
based detection and classification of melanoma skin cancer using
this field of research. hybrid texture features. Sensors International, 2, 100128.
[4]. Kumar, M., Ray, S., & Yadav, D. K. (2021). A survey on moving
object detection in video using a moving camera for smart
V. CONCLUSION surveillance system. In Proceedings of International Conference on
Machine Intelligence and Data Science Applications: MIDAS
A proposed hybrid model of skin cancer detection 2020 (pp. 241-253). Springer Singapore.
integrated binary classifier, Residual Network, and CNN [5]. Inthiyaz, S., Altahan, B. R., Ahammad, S. H., Rajesh, V., Kalangi,
which shows good results on dermoscopic image. This R. R., Smirani, L. K., ... & Rashed, A. N. Z. (2023). Skin disease
detection using deep learning. Advances in Engineering
robust integrated model helps in the accurate detection
Software, 175, 103361.
and classification of skin cancer in the early stage

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1121

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of 2nd International Conference on Sustainable Computing and Smart Systems (ICSCSS 2024)
IEEE Xplore Part Number: CFP24DJ3-ART; ISBN: 979-8-3503-7999-0

[6]. Ghaffar Nia, N., Kaplanoglu, E., & Nasab, A. (2023). Evaluation of Smart Cities. Journal of Information Technology
artificial intelligence techniques in disease diagnosis and Management, 14(Special Issue: Security and Resource Management
prediction. Discover Artificial Intelligence, 3(1), 5. challenges for Internet of Things), 118-131.
[7]. Pilania, U., Kumar, M., Singh, S., & Adhalkha, P. (2023, [14]. Pilania, U., Diwakar, C., Arora, K., & Chaudhary, S. (2022,
December). A Deep Ensemble Learning Network-based Approach September). An Optimized Hybrid Approach to Detect Cataract.
to Detect Neurodegenerative Diseases. In 2023 2nd International In 2022 IEEE Global Conference on Computing, Power and
Conference on Automation, Computing and Renewable Systems Communication Technologies (GlobConPT) (pp. 1-5). IEEE.
(ICACRS) (pp. 1033-1038). IEEE. [15]. Pilania, U., Tanwar, R., & Nandal, N. (2022). Exploring face
[9]. Nambisan, A. K., Maurya, A., Lama, N., Phan, T., Patel, G., Miller, detection and recognition in steganography. Unleashing the art of
K., ... & Stoecker, W. V. (2023). Improving automatic melanoma digital forensics, 181-214.
diagnosis using deep learning-based segmentation of irregular [16]. Vidya, M., & Karki, M. V. (2020, July). Skin cancer detection
networks. Cancers, 15(4), 1259. using machine learning techniques. In 2020 IEEE international
[10]. Pilania, U., Tanwar, R., & Nandal, N. (2022). Exploring face conference on electronics, computing and communication
detection and recognition in steganography. Unleashing the art of technologies (CONECCT) (pp. 1-5). IEEE.
digital forensics, 181-214. [17]. Monika, M. K., Vignesh, N. A., Kumari, C. U., Kumar, M. N. V. S.
[11]. Orhan, H., & YAVŞAN, E. (2023). Artificial intelligence-assisted S., & Lydia, E. L. (2020). Skin cancer detection and classification
detection model for melanoma diagnosis using deep learning using machine learning. Materials Today: Proceedings, 33, 4266-
techniques. Mathematical Modelling and Numerical Simulation with 4270.
Applications, 3(2), 159-169. [18]. Daghrir, J., Tlig, L., Bouchouicha, M., & Sayadi, M. (2020,
[12]. Kumar, M., Yadav, D. K., Ray, S., & Tanwar, R. (2024). Handling September). Melanoma skin cancer detection using deep learning
illumination variation for motion detection in video through and classical machine learning techniques: A hybrid approach.
intelligent method: An application for smart surveillance In 2020 5th international conference on advanced technologies for
system. Multimedia Tools and Applications, 83(10), 29139-29157. signal and image processing (ATSIP) (pp. 1-5). IEEE.
[13]. Kumar, M., Ray, S., & Yadav, D. K. (2022). Moving Vehicles
Detection and Tracking on Highways and Transportation System for

979-8-3503-7999-0/24/$31.00 ©2024 IEEE 1122

Authorized licensed use limited to: MULTIMEDIA UNIVERSITY. Downloaded on January 17,2025 at 06:36:32 UTC from IEEE Xplore. Restrictions apply.