Journal of Robotics and Control (JRC) Volume 7.
Issue 3, 2026 ISSN: 2715-5072.
DOI: 10.
18196/jrc.
Lightweight Hybrid Feature Extraction for RealTime Skin Lesion Classification on HAM10000 and MSK10000 Fallah H.
Najjar 1, 2, a, *.
Farhan Mohamed 1, 3, b, *.
Ali Hilal Ali 4, c.
Vei Siang Chan 1, 3, d.
Mohd Shafry Mohd Rahim 1, 3, e.
Alfian Ma'arif 5, f.
Karrar A.
Kadhim 1, 6, g Department of Emergent Computing.
Faculty of Computing.
Universiti Teknologi Malaysia, 81310 UTM Johor Bahru.
Johor.
Malaysia Department of Computer Networks and Software Techniques.
Technical Institute of Najaf.
Al-Furat Al-Awsat Technical University, 54001 Najaf.
Iraq Media and Game Innovation Center of Excellence.
Institute of Human Centered Engineering.
Universiti Teknologi Malaysia, 81310 UTM Johor Bahru.
Johor.
Malaysia Electronic and Communications Department.
Faculty of Engineering.
University of Kufa.
PO.
Box 21.
Najaf, 54001.
Iraq Universitas Ahmad Dahlan.
Yogyakarta.
Indonesia Computer Techniques Engineering Department.
Faculty of Information Technology.
Imam JaAoafar Al-Sadiq University, 10001.
Baghdad.
Iraq Email: a fallahnajjar@atu.
iq, b farhan@utm.
my, c alih.
alathari@uokufa.
iq, d vschan2@live.
my, e shafry@utm.
alfianmaarif@ee.
id, g karrar.
abdulameer@ijsu.
*Corresponding Authors AbstractAi Skin cancer is one of the most prevalent, and in some cases deadly, cancers globally, underscoring the importance of early and precise diagnosis upon which successful disease management is dependent.
Conventional diagnosis procedures are time-consuming and prone to varying routines requiring automated and consistent classification systems.
this paper, we propose a hybrid feature extraction method for binary and multi-class skin-lesion classification by integrating Local Binary Pattern (LBP) and Gray Level Run Length Matrix (GLRLM).
The approach is benchmarked on two public dermoscopic image datasets (HAM10000, seven classes and MSK10000, binary classificatio.
Twenty-one features are calculated for each image .
LBP and 11 GLRLM), and their classification is done using five machine learning algorithms: kNearest Neighbor .
NN).
Support Vector Machine (SVM).
Decision Tree (DT).
Random Forest (RF) and AdaBoost.
The proposed system achieves a maximum 99.
52% accuracy and a 25% F1-score in the binary classification, outperforming many deep learning based benchmarks.
multi-class problems, it reaches 97.
63% accuracy and presents high scores in precision, sensitivity, and specificity.
Crossvalidation verifies the stability of the model.
Feature extraction requires only 0.
011 seconds per image, revealing computational Analysis of feature importance shows a consistent contribution by all the features extracted.
The technique exhibits high performance, minimal computational burden, and is deployable in a real-time clinical setup.
This method provides a convenient solution for early and accurate skin lesion classification in both clinical and resource-constrained KeywordsAi Skin Cancer.
Early Diagnosis.
HAM10000.
MSK10000.
Lightweight Classification.
INTRODUCTION
Skin cancer is a prevalent cancer in the world .
, which has an increasing occurrence due to, for instance, augmented exposure to UV radiation and aging photo-sensitive populations .
, .
Early and exact diagnosis is very necessary, because treatment in time may essentially enhance the survival rate of patients .
, .
Dermatologists commonly perform their diagnosis based on naked eye and dermoscopic examination to detect malignant skin lesions .
However, this approach is time-consuming, subjective in nature, and it inevitably depends on inter-observer variance .
, .
Therefore, there has been a high demand for automatic diagnostic techniques based on medical image analysis and machine learning.
Texture-based feature extraction methods are one among such attempts, which are found to be effective in capturing the structural patterns of skin lesions .
Local Binary Pattern (LBP) and Gray Level Run Length Matrix (GLRLM) are a pair of commonly-used algorithms that measure local texture and gray-scale intensity distributions, respectively .
Both have been applied separately in the context of skin lesion analysis.
However, their joint application in the form of a hybrid feature extraction technique for binary and multiclass classification is less investigated in the literature .
This discrepancy becomes even clearer when we focus on lightweight and interpretable methods, appropriate for real-time exploitation on low-power devices.
Most of the existing works tend to focus on accurate deep learning models at the cost of complex computations and a lack of interpretability .
, .
In contrast to the existing prior art, in this study, a LBP-GLRLM hybrid feature extraction pipeline is proposed.
It is expressive for achieving a balance between performance and time efficiency.
assess our approach on two widely used public datasets: the Melanoma Skin Cancer Dataset of 10000 Images (MSK10.
, binary classification, benign vs malignant, and the Human Against Machine dataset (HAM10.
, multi-class, addressing seven diagnostic categories.
Journal Web site: http://journal.
id/index.
php/jrc Journal Email: jrc@umy.
Journal of Robotics and Control (JRC) ISSN: 2715-5072 Preliminary class imbalance, resolution variations, and noise in these data are rectified by data augmentation, normalization, and stratified sampling .
etailed in the methodology sectio.
Our approach combines, implements, and evaluates a feature extraction method based on LBP and GLRLM for skin cancer classification.
Our approach distinguishes itself by going beyond traditional LBP feature extraction and harnessing the potential of LBP matrix features.
This innovative approach enhances the classification accuracy and provides a more comprehensive analysis of skin lesion textures, making it a valuable tool in dermatology and medical image analysis, as demonstrated in the results This innovative method extracts 21 features, ten from LBP and eleven from GLRLM, providing a comprehensive feature set that effectively captures skin lesions' texture and gray-scale information.
Nevertheless, for the evaluation of the proposed method, five machine learning classifiers, namely k-Nearest Neighbor .
NN).
Support Vector Machine (SVM).
Decision Tree (DT).
Random Forest (RF), and AdaBoost, have been selected to perform the classification of the extracted features.
The contributions of our method are summarized as .
Incorporating data augmentation by rotating each image at three angles .
A, 135A, and 235A).
Extracted 21 features from skin cancer datasets using ten features from the LBP and eleven features from the GLRLM method to increase classification .
Suggested five different classifiers to classify the datasets into their corresponding classes.
Numerical results on the MSK10000 and HAM10000 datasets show that our method outperforms the stateof-the-art approaches concerning speed, accuracy, and The remaining sections of this paper are organized as follows: Section II will provide a thorough review of relevant literature on skin lesion classification.
Section i will expound upon the proposed feature extraction method, elucidating its intricacies.
Section IV will illustrate the details of the datasets.
Section V will describe the experimental setup and present the results and performance analysis.
Finally.
Section VI will conclude the paper, underscoring its contributions and highlighting potential avenues for future II.
RELATED WORKS
Over the past two decades, extensive research has been conducted in Computer-Aided Diagnosis (CAD) systems to detect and identify skin cancer.
Various approaches have been explored, from classical artificial intelligence to machine learning and deep learning techniques .
, .
Recent studies have mainly focused on employing machine learning and deep learning methods to automate melanoma detection .
Consequently, numerous methodologies and strategies have been proposed in published works to facilitate CAD For instance.
Ain et al.
conducted several studies using Genetic Programming (GP) as a feature selection method in skin cancer detection.
In one study, they used GP to develop a skin cancer detection classifier by combining domain-specific features from dermatologists with LBP features from dermoscopic images.
Another study introduced a method that used LBP for feature extraction.
GP for feature selection, and feature engineering to enhance classification accuracy, incorporating gray-scale and color features to distinguish between benign and malignant skin They also developed a two-stage methodology for feature selection and construction using GP and LBP features from dermoscopy images .
Additionally, they created an innovative classification method using multi-tree GP to effectively distinguish between ten different skin cancer classes based on lesion images, considering both local and global information.
The choice of feature extraction techniques and their integration was crucial in improving skin cancer classification performance .
lkin et al.
proposed an accurate melanoma detection approach employing a hybrid classification system consisting of the SVM classifier and a bacterial colony optimization However, this method exhibited two limitations.
Firstly, the algorithmAos performance deteriorated when processing images with noise, such as those containing hair or Secondly, when the affected area extended beyond the boundaries of the image, the reliability of the classifierAos features diminished.
Dhivyaa et al.
explored the utilization of DT and random forest classifiers for skin lesion classification.
The proposed method underwent testing on the HAM10000 and ISIC 2017 datasets.
Regrettably, when confronted with a large dataset, the method experiences limitations and is susceptible to image noise.
In addition, several deep learning studies were proposed to diagnose skin cancer, especially with the HAM10000 for example.
Shetty et al.
presented a new Convolutional Neural Network (CNN) for Skin cancer they used 150 epochs in the training process.
Even with 150 epochs, the highest accuracy values were 18% and 86.
43% for training and training sets.
The difference in accuracy between the training and testing sets can be a sign of overfitting.
In this case, the model may have learned the specific patterns in the training data but needs help generalizing them to new data.
Popescu et al.
introduced a multi-neural network for Skin cancer classification using nine networks with a collective intelligence block.
Still, with the complexity of their proposed method, the highest accuracy was obtained at The main limitation of their proposed method is its high computational complexity.
The nine CNNs need to be trained individually, which is computationally expensive.
Additionally, the collective intelligence block also requires a significant amount of computation.
As a result, the proposed method is unsuitable for deployment on portable Another area for improvement of the proposed Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 method is that it was only evaluated on the HAM10000 It is conceivable that the methodAos performance on alternative datasets may not be as satisfactory.
Shahin et al.
proposed an algorithm for skin cancer classification that combined two CNNs: Inception V3 and ResNet-50.
The two networks were first trained individually, and then their outputs were combined using an ensemble learning method.
Their proposed method achieved an accuracy of 89.
90% on the HAM10000 dataset.
Finally.
Srinivasu et al.
proposed using the MobileNet-V2 with LSTM for skin cancer classification.
The MobileNet-V2 is a CNN designed to be efficient, while the LSTM is a recurrent neural network designed to learn longterm dependencies.
Their proposed method achieved an accuracy of 88.
90% on the HAM10000 dataset.
The main limitation of their proposed method is that it needs to be sufficiently random.
The input images are not randomly cropped or augmented, and the outputs of the MobileNet-V2 are not randomly shuffled, which means that the model cannot learn all possible patterns in the data.
Jadhav et al.
proposed a CNN to extract features of skin lesions, and an SVM was used to classify those features.
The CNN was used without pre-processing for feature extraction, eliminating the need for hand-crafted features.
Their proposed method achieved an accuracy of 87.
25% on the HAM10000 dataset.
The main limitation of their proposed method is that it is highly memory-intensive.
The CNN needs to store all of the activations of the hidden layers, which can be a significant amount of data, making their proposed method challenging to deploy on devices with limited memory, such as mobile phones.
Srinivasu et al.
proposed a Heuristic Approach for Real-Time Image Segmentation (HARIS) for skin cancer The HARIS algorithm is a two-phase algorithm that first segments the image into regions and then classifies each region.
The segmentation phase uses a heuristic approach to identify the optimal number of regions in the image.
The classification phase uses the SVM classifier to classify each region.
Their proposed method achieved an accuracy of 84.
80% on the HAM10000 dataset.
Zhang et al.
proposed a multi-class lesion classification strategy combining the Feature Pyramid Network with Resnet-50.
Their proposed approach was to train the CNN on all 1279 images from the ISBI 2016.
PH2, and ISIC 2017 datasets, and then they tested it on the HAM10000 dataset.
However, they stated that the obtained results for the test set of the HAM10000 dataset were 86.
5%, 85.
5%, 87.
0%, and 86.
for accuracy, sensitivity, precision, and F1-score.
The method could be more robust to noise and variations in image quality.
The Feature Pyramid Network is relatively shallow, meaning it may need help to learn to classify noisy images or have variations in image quality.
Alli et al.
propose an enhanced data augmentation model to detect melanoma skin cancer efficiently.
Utilizing dermoscopy images from the publicly available PH2 dataset, they apply their proposed data augmentation technique to create a new skin melanoma dataset.
They achieved an accuracy of 95% in binary classification, while multi-class classification yielded only 66%.
However, their method could have been multi-class Bian et al.
proposed a deep learning-based technique for skin lesion classification that achieved promising results on three publicly available datasets.
On the ISIC-2016 dataset, their model achieved an accuracy of 84.
7%, a sensitivity of 8%, a specificity of 88.
1%, and an average precision of On the ISIC-2017 dataset, their model achieved an accuracy of 86.
2%, a sensitivity of 62.
2%, and a specificity of On the HAM10000 dataset, their model achieved a sensitivity of 76.
9% and a specificity of 96.
Kadry et al.
used a pre-trained VGG-SegNet algorithm to extract a skin melanoma region from the ISIC2016 dataset image with 224x224x3 pixels dimensions.
Crucial performance parameters were determined after contrasting the extracted segmented with the ground truth.
However, their proposed technique was concerned with segmentation only.
However, they achieved a segmentation accuracy of 97.
Nawaz et al.
presented a framework for melanoma segmentation that achieved a segmentation accuracy of 99% for both the ISIC-2017 and ISIC-2018 datasets.
However, their method does not efficiently diagnose melanoma moles under extreme intensity variations.
Their method is based on the UNet model, a CNN designed to work well on images with a relatively uniform intensity distribution.
However, images of melanoma moles can have a wide range of intensity variations, making it difficult for the UNet model to segment them accurately.
In our previous works .
, a Deep Learning (DL) approach was put forth for the early detection of melanoma.
This strategy incorporated a CNN network architecture featuring five convolution layers, five pooling layers, a fully connected layer, an input layer, and an output layer.
Utilizing the MSK10000 dataset, the system attained an accuracy level of 91%.
Yaman et al.
proposed a hybrid method that fused deep feature extraction with an enhanced variant of the minimum redundancy maximum relevance .
RMR) feature selection method called ImRMR.
Five pre-trained deep learning models were employed as feature generators to derive 1000 features from each model, totaling 5000 features.
The ImRMR feature selection procedure was employed iteratively to automatically pick the most pertinent features, eliminating the requirement for a trial-and-error strategy.
Subsequently, these chosen features were input into a Deep Neural Network (DNN) for classification.
Their proposed model, presented in the study, achieved a remarkable accuracy rate of 96.
58% when evaluated on the HAM10000 skin cancer image dataset.
Sharma et al.
investigate using machine learning and deep learning models for classifying benign and malignant skin lesions.
The authors used four machine learning models:
SVM.
Naive Bayes (NB), kNN, and Neural Networks (NN).
They also used a pre-trained SqueezeNet model for feature extraction from the MSK10000 dataset.
Their study showed that the NN model outperformed the other three models in terms of accuracy.
F1 measure, recall, precision.
AUC, and ROC.
The NN model achieved an accuracy of 88.
2%, which is significantly higher than the accuracies of the other three Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 models (SVM: 82.
NB: 80.
8%, kNN: 79.
2%).
Finally.
Balaha et al.
proposed another deep learning approach for skin cancer detection, segmentation, and classification.
Using the Harris Hawks optimization algorithm.
The authors proposed optimizing five different pre-trained CNN methods:
MobileNet.
DenseNet201.
DenseNet169.
VGG19, and VGG16.
Their proposed method was applied to two datasets.
MSK10000 and ISIC.
However, their proposed method had some restrictions.
Time-consuming is the most notable advantage because it takes the longest time to train a i.
The LBP and GLRLM feature extraction techniques were also computationally efficient.
Both techniques are known for their low computational cost and quick execution times, making them suitable for real-time applications and implementation on portable devices with limited computational resources.
By focusing on these traditional classifiers and efficient feature extraction methods, our approach ensures a balance between performance and complexity, making it feasible for deployment in diverse settings, including edge devices with constrained computational capabilities.
Local Binary Pattern (LBP) Ojala et al.
were the ones who initially presented the LBP operator.
This operator works with a pixel's immediate eight neighbors, with the value of the center pixel serving as a threshold for the operation.
If a neighbor pixel has the same gray value as the center pixel or a higher gray value than the center pixel, then one is assigned to that pixel.
If not, then it gets a zero.
Finally, the eight ones or zeros are concatenated to a binary code, resulting in the production of the LBP code for the center pixel, as shown in Fig.
The LBP operator was later modified to accommodate a broader range of neighborhood sizes.
In this case, we draw a circle with the center at pixel and radius R.
The focal pixel's value is compared with readings taken from P sampling points around the circle's circumference.
Any radius (R) and a different number of pixels (P) can be used to obtain the values of all nearby sampling points.
The (P.
R) notation is used to denote local neighborhoods.
Three neighbor sets for varying values of P and R are shown in Fig.
However, the modified LBP is also called extended or multiscale LBP.
MACHINE LEARNING WORKFLOW
Whereas our study primarily focuses on traditional machine learning classifiers like kNN.
SVM.
DT.
RF, and AdaBoost and not on neural network architectures, it is important to discuss the complexity of our approach in terms of computational cost and ease of implementation.
The kNN algorithm inherently has a low model complexity as it does not involve any training phase.
The primary computational cost is during the prediction phase, comparing a test instance with all training instances.
However, considering the dataset sizes used in our experiments, the computation was still manageable and efficient.
On the other hand, the SVM classifier is slightly more complex than kNN.
It involves finding the optimal hyperplane that distinctly classifies the data points into different classes.
DT involves creating a treelike graph of decisions.
The complexity of DT is relatively low and is determined by the depth of the tree.
We ensured the tree depth was optimal to avoid overfitting while maintaining computational efficiency.
Threshold Ie Fig.
LBP operator, binary: 11101001 Ie decimal: 233, blue: start, red: end, gray: focal pixel.
Fig.
Extended or multiscale LBP operator .
Further, the LBP feature extraction operator is denoted in Eq.
ycE yayaAycE.
cE,ycI) = Oc ycI.
aycE Oe yayca )ycn ycn=1 Where ycE represents the pixel number in the neighborhood, ycI.
= 1 ycnyce yca Ou 0, ycuycEayceycycycnycyce 0.
However, the LBP algorithm generates feature matrices for each image, containing ten image features .
TABLE I.
LBP FEATURES
Features Mathematical expression Eq.
ycuOe1 ycoOe1 Contrast .
, .
Oc Oc .
c Oe yc.
2 y ycnyco.
c, yc.
yc=0 yca=0 ycuOe1 ycoOe1 Energy .
, .
oc Oc ycnyco.
c, yc.
2 yc=0 yca=0 ycuOe1 ycoOe1 Entropy .
, .
Correlation .
Dissimilarity .
Joint Variance .
, .
Joint Average .
Cluster Prominence .
Cluster Shade .
Auto Correlation .
, .
Oc Oc ycnyco.
c, yc.
y log.
c, yc.
) yc=0 yca=0 ycuOe1 ycoOe1 Oc Oc .
c Oe yuNyc .
ca Oe yuNyc .
) ycnyco.
c, yc.
yuayc2 Oc Oc .
c Oe yc.
c, yc.
yc=0 yca=0 ycuOe1 ycoOe1 yc=0 yca=0 ycuOe1 ycoOe1 Oc Oc .
c Oe yuNyca )2 ycnyco.
c, yc.
yc=0 yca=0 ycuOe1 ycoOe1 Oc Oc yc ycnyco.
c, yc.
ycuOe1 ycoOe1 Oc Oc .
c Oe yca Oe yuNycyycu yuNycyyc } ycnyco.
c, yc.
yc=0 yca=0 ycuOe1 ycoOe1 Oc Oc .
c Oe yca Oe yuNycyycu yuNycyyc } ycnyco.
c, yc.
yc=0 yca=0 yc=0 yca=0 ycuOe1 ycoOe1 Oc Oc yc yca ycnyco.
c, yc.
yc=0 yca=0 Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 Table I and Table II delineate the features derived from applying LBP .
, 51-.
and GLRLM .
feature extraction methodologies within the skin cancer detection system, respectively .
, 51-.
Gray Level Run Length Matrix (GLRLM) GLRLM counts the number of consecutive pixels with the same gray level value.
In the GLRLM ycE.
c, yc.
yuE), the notation .
c, yc.
2 describes the frequency with which runs of gray level yc and length yco appear in the image at each angle yuE.
Eq.
c, yc.
yuE) y OcOc ycAyc .
uE) yc2 TABLE i.
HYPERPARAMETERS SUMMARIZATION
y Oc Oc yco2 y ycE.
c, yc.
yuE) ycAyc .
uE) y Oc(Oc yuU.
c, yc.
yuE))2 ycAyc .
uE) TABLE II.
GLRLM FEATURES
Run Percentage (RP) .
Mathematical expression ycAyci ycAyc yc=1 yco=1 ycAyci ycAyc Classifier SVM yc=1 yco=1 ycAyci ycAyc yc=1 yco=1 ycAyc ycAyci y Oc(Oc ycE.
c, yc.
yuE))2 ycAyc .
uE) yco=1 yc=1 ycAyci ycAyc OcOc yc=1 yco=1 Low Gray Run Emphasis (LGRE) .
High Gray Run Emphasis (HGRE) .
Short Run Low Gray Emphasis (SRLGE) .
Short Run High Gray Emphasis (SRHGE) .
Long Run Low Gray Emphasis (LRLGE) .
Long Run High Gray Emphasis (LRHGE) .
step in our proposed method, aimed at enhancing the diversity and robustness of the dataset.
In this stage, we employ a rotation-based augmentation strategy to generate additional training samples from the original dataset.
The data rotation transformations expand the training data and help the model become more invariant to different orientations and For each input image, we apply rotation transformations at three specific angles: 45A, 135A, and 225A.
This strategy effectively simulates different orientations of the objects within the images, capturing variations that may be encountered in real-world scenarios.
Introducing these augmented samples into the training set makes our model more adept at recognizing objects from different angles and can better generalize to unforeseen orientations during the testing phase.
However.
Table i comprehensively summarizes the hyperparameters and their corresponding values for the SVM, kNN.
DT.
RF, and AdaBoost classifiers.
The GLRLM is more concerned with a group of adjacent, gray-level pixels running in the same direction.
GLRLM is typically calculated in four directions, with a run-length histogram for each .
, .
Features Short Run Emphasis (SHE) .
Long Run Emphasis (LHE) .
Gray-Level Nonuniformity (GLN) .
Run Length Nonuniformity (RLN) .
y ycE.
c, yc.
yuE) ycAycy ycAyci ycAyc ycE.
c, yc.
yuE) y OcOc ycAyc .
uE) yc2 yc=1 yco=1 ycAyci ycAyc y Oc Oc yc 2 y ycE.
c, yc.
yuE) ycAyc .
uE) yc=1 yco=1 ycAyci ycAyc ycE.
c, yc.
yuE) y OcOc 2 ycAyc .
uE) yc y yco2 yc=1 yco=1 ycAyci ycAyc ycE.
c, yc.
yuE) y yc y OcOc ycAyc .
uE) yco2 AdaBoost yc=1 yco=1 ycAyci ycAyc ycE.
c, yc.
yuE) y yco2 y OcOc ycAyc .
uE) yc2 Oc Oc ycE.
c, yc.
yuE) y yc 2 y yco2 ycAyc .
uE) yc=1 yco=1 ycAyci ycAyc yc=1 yco=1 where: ycAyci represents the number of discrete intensity values in the image, ycAyc represents the number of discrete run lengths, ycAycy represents the number of voxels, ycAyc .
uE) represents the number of runs along an angle yuE, and it is equal ycAyci ycAyc to Ocycn=1 Ocyc=1 ycE.
c, yc.
yuE), and 1 O ycAyc .
uE) O ycAycy , and ycE.
cn, y.
yuE) represents the normalized run length matrix yuU.
c, yc.
yuE) = ycE.
c, yco .
uE ) ycAyc .
uE) Proposed method Our proposed method comprises four stages: dataset augmentation, feature extraction, classification, and Data augmentation is a crucial pre-processing Hyperparameter KernelFunction Gaussian Value KernelScale BoxConstraint Standardize TRUE Cityblock Distance NumNeighbors DistanceWeight Equal Standardize TRUE SplitCriterion Deviance MaxNumSplits Surrogate Off MaxNumSplits Method Bag NumLearningCycles MaxNumSplits Method AdaBoostM2 NumLearningCycles Consequently, the feature extraction starts by extracting ten features using the LBP method and 11 features using the GLRLM method.
We will elaborate on the details of these techniques in the following subsection.
The feature extraction stage results in a feature dataset containing .
when applied to the MSK10000 dataset and .
for the HAM10000 dataset.
In order to ensure a fair comparison and to maintain consistency with our previous work, the MSK10000 dataset was randomly split into two parts: 85% for training and 15% for testing, following the division described in the referenced research paper .
In contrast, most cutting-edge techniques employed an 80% training and 10% testing split for the HAM10000 dataset.
Furthermore, in order to address both overfitting and bias reduction, we employed a technique known as 10-fold crossvalidation.
We utilized it to evaluate the performance of a model while mitigating the risks associated with overfitting Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 and bias.
This approach divides the dataset into ten equally sized subsets or "folds" .
We then iteratively train the model using nine folds and evaluate its performance on the remaining fold.
This process is repeated ten times, with each fold being the validation set once.
Three classifiers, kNN.
SVM, and DT, are trained on the training sets alongside the 10-fold cross-validation.
The performance of each classifier is evaluated using six metrics: Accuracy (Ac.
Precision (Pr.
Sensitivity (Se.
F1-score (F.
Specificity (Sp.
, and Matthews Correlation Coefficient (MCC).
Fig.
3 depicts the workflow stages of this TABLE IV.
MSK10000 AND HAM10000 DATASETS CLASS
DISTRIBUTION
Dataset
MSK10000
HAM10000
Classes benign .
0 image.
5 image.
melanocytic Nevi (NV) .
5 melanoma (MEL) .
3 image.
benign keratosis-like lesions (BKL) .
9 image.
basal cell carcinoma (BCC) .
actinic keratoses and intraepithelial carcinoma (AKIEC) .
vascular lesions (VASC) .
dermatofibroma (DF) .
Total images .
It is clear from Table IV that the two datasets are different in terms of the number of classes.
The MSK10000 has two classes .
therefore, it is a binary classification problem, whereas the HAM10000 has seven classes .
We will report the results of each dataset in the following sections.
EXPERIMENTAL RESULTS
Training and testing of the machine learning algorithms were performed on a local PC with Windows 11.
Intel Core i7-11800H .
30 GHz, 16 CPU), 16 GB DDR4 RAM, and NVIDIA GeForce RTX 3050 TI GPU and MATLAB (R2022.
The following section will report and discuss our Fig.
Proposed workflow.
Finally, the ROC curve and AUC of the system were calculated and plotted.
the ROC curve can be drawn using either class probabilities or predicted class labels.
However, using predicted class labels is generally considered more informative .
IV.
DATASETS
This section presents detailed information about the datasets employed in training the machine learning algorithms for skin cancer classification.
Specifically, we utilized two datasets with distinct characteristics and significance in the field.
The first dataset.
MSK10000 .
, is a relatively new addition to dermatology and skin cancer research, published in 2022.
Its novelty brings fresh perspectives and challenges to skin cancer classification.
On the other hand, the second dataset.
HAM10000 .
, is a well-established and widely recognized dataset within the research community.
It has been extensively referenced and used in numerous studies, making it a benchmark for evaluating the performance of skin cancer classification Table IV provides a comprehensive summary of the pertinent characteristics associated with each dataset, consolidating the necessary details for easy reference.
By utilizing both MSK10000 and HAM10000 in our experiments, we aim to leverage the strengths of both datasets to develop a robust and effective skin cancer classification model that can contribute to the ongoing efforts to improve diagnostic accuracy in dermatology.
Binary Classification In Table V, we report the experimental results of our proposed method as applied to the MSK10000 dataset.
will compare our results to the only three research efforts in the literature we found .
, .
, and .
In addition, kNN.
SVM.
DT.
RF, and AdaBoost classifiers were chosen to evaluate the effectiveness of our approach.
All classifiers were tested using 15% of the dataset for testing and 85% for Classifier efficacy is determined by comparing the classifierAos predicted and actual labels.
The efficacy of our proposed method was evaluated using the Acc.
Pre.
Spe.
Sen.
F1, and MCC metrics.
TABLE V.
BINARY CLASSIFICATION RESULTS USING FIVE CLASSIFIERS
Metrics Acc
Pre
Sen
Spe
MCC
(%)
SVM
(%)
(%)
(%)
AdaBoost
(%)
Table VI compares the proposed algorithmAos performance with the state-of-the-art approaches to the MSK10000 dataset.
Table VI shows the effectiveness of our feature extraction method in accurately classifying the skin cancer image Five different machine learning algorithms are used for classification: kNN.
SVM.
DT.
RF, and AdaBoost.
They all show results of at least 98%, suggesting outstanding Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 TABLE VI.
COMPARISON OF RESULTS WITH STATE-OF-THE-ART
METHODS REGARDING THE MSK10000 DATASET
Methods Waheed et al.
Sharma et al.
DenseNet169 .
VGG19 .
VGG16 .
MobileNet .
DenseNet201 .
Proposed .
NN) Proposed (SVM) Proposed (DT) Proposed (RF) Proposed (AdaBoos.
Acc (%) Pre (%) Sen (%) Spe (%) Table VI shows a comparison to the state-of-the-art methods in the literature.
Although the approaches in the literature are primarily deep learning ones, we have shown how a more straightforward approach can be more effective with a better-suited feature extraction routine.
Additionally.
Fig.
4 illustrates the confusion matrix and ROC curve.
Finally, the results indicate the absence of bias in the proposed method.
Fig.
Binary class confusion matrices and ROC curves: .
kNN, .
SVM, .
DT, .
RF, and .
AdaBoost.
Multi-class Classification In Table VII, we report our experimental results of the proposed method when applied to the HAM10000 with the same performance metrics described in previous sections.
TABLE VII.
MULTI-CLASS CLASSIFICATION RESULTS USING FIVE
CLASSIFIERS REGARDING THE HAM10000 DATASET
Metrics Acc
Pre
Sen
Spe
MCC
KNN
(%)
SVM
(%)
DT (%)
RF (%)
AdaBoost
(%)
Once more, the proposed feature extraction routine has performed better in all major machine learning algorithms:
SVM, kNN.
DT.
RF, and AdaBoost.
Although the metric values are lower than the previous binary classification, we will show in Table Vi that it is better than all approaches in the literature.
TABLE Vi.
COMPARISON OF RESULTS WITH STATE-OF-THE-ART
METHODS REGARDING THE HAM10000 DATASET
Methods Shetty et al.
Popescu et al.
Shahin et al.
Srinivasu et al.
Synmez et al.
Proposed .
NN) Proposed (SVM) Proposed (DT) Proposed (RF) Proposed (AdaBoos.
Acc (%) Pre (%) Sen (%) Spe (%) The highest accuracy is 97.
63%, registered with the SVM algorithms equipped with our feature extraction method.
It is higher by at least 10% than those in the literature.
This clearly shows the versatility of the feature extraction routine even on the multi-class dataset and proves it is a strong candidate for low-computation applications such as edge devices.
Additionally.
Fig.
5 introduces the multi-class confusion matrix of the classifierAos test results and the system's ROC curve and AUC.
Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 Therefore, the results showed that all features were significant in the classification task, with minor variances in their importance scores.
Although the features Contrast.
ClusterShade.
SRE.
SRLGE, and SRHGE were ranked with marginally higher values, no single feature can be largely distinguished in the prediction process.
However, this balanced contribution underscores the effectiveness of our hybrid feature extraction strategy.
It confirms that the selected features collectively capture a wide range of texture and gray-scale information critical for robust skin lesion classification.
The feature importance plot in Fig.
6 illustrates the distribution of these values, further demonstrating that each feature adds value to the classification process.
Fig.
Feature Importance Scores Calculated Using the RF Classifier.
Cross-Validation Classification This section presents the results obtained from employing a technique known as 10-fold cross-validation to address the challenges of overfitting and bias.
Table IX presents the binary classification results obtained from three different classifiers applied to the MSK10000 dataset using a 10-fold cross-validation technique.
Table XI presents the results of multi-class classification obtained from three different TABLE IX.
BINARY CLASSIFICATION RESULTS USING THREE CLASSIFIERS
REGARDING THE MSK10000 DATASET
Metrics Acc
Pre
Sen
Spe
MCC
Fig.
Multi-class class confusion matrices and ROC curves: .
kNN, .
SVM, .
DT, .
RF, and .
AdaBoost.
In addition, to make our classification model more interpretable and corroborate the validity of our high accuracy results, we used the feature importance function of the RF classifier.
This analysis assessed the relative impact of each extracted feature, ten from LBP and eleven from GLRLM, on the final classification outcome.
KNN (%)
SVM (%)
DT (%)
The results demonstrate that all three classifiers performed remarkably well on the MSK10000 dataset.
The SVM classifier generally exhibited the highest scores across most metrics, indicating its superior performance in accurately classifying instances in this dataset.
However, the kNN and DT classifiers also achieved high accuracy and firm performance in various evaluation measures, making them viable alternatives depending on specific requirements and Furthermore, the Mean A SD of performance of each classifier evaluated by 10-fold cross-validation is shown in Table X.
The average F1-score was the highest for the SVM Fallah H.
Najjar.
Lightweight Hybrid Feature Extraction for Real-Time Skin Lesion Classification on HAM10000 and MSK10000 Journal of Robotics and Control (JRC) ISSN: 2715-5072 .
52% A 0.
, closely followed by the DT and kNN TABLE X.
BINARY CLASSIFICATION RESULTS USING THREE CLASSIFIERS
REGARDING THE MSK10000 DATASET
Metrics Acc
MCC
SVM (%)
48 A 0.
52 A 0.
96 A 0.
kNN (%) 22 A 0.
23 A 0.
45 A 0.
DT (%)
67 A 0.
68 A 0.
35 A 0.
the case of the HAM10000 dataset, we get about 0.
seconds for each image.
Hence, this time represents the time required for feature extraction, including data augmentation.
By dividing this number by 4, the real-time of extracting 21 features from one image is 0.
011 and 0.
0301 seconds for MSK10000 and HAM1000, respectively.
Given the type of hardware used for conducting this experiment, our approach can be easily implemented on a budget portable hardware.
VI.
We performed a Wilcoxon signed-rank test for SVM vs.
DenseNet169 .
that achieved 97.
08% accuracy on different scenarios.
Even though our approach obtained an accuracy improvement of 2.
4%, this observation was not significant .
> 0.
as a result of the protocol variation.
is important to stress that our claimed superiority is situational and not an absolute one.
TABLE XI.
MULTI-CLASS CLASSIFICATION DATASET
Metrics Acc
Pre
Sen
Spe
MCC
KNN (%)
SVM (%)
DT (%)
The results indicate that the performance of all three classifiers was commendable on the HAM10000 dataset.
The SVM classifier consistently outperformed the other classifiers across most metrics, underscoring its ability to classify instances across multiple classes in this dataset Nevertheless, it is worth noting that all the kNN and DT classifiers also attained high levels of accuracy and exhibited robust performance in various evaluation measures, presenting viable alternatives that can be considered based on specific requirements and preferences.
Feature Extraction Period This section will report on the feature extraction required time for each dataset in this study.
Twenty-one features were extracted from each dataset.
The number of samples of the MSK10000 is 42420, and 40060 samples for the HAM10000 TABLE XII.
FEATURE EXTRACTION PERIOD
Dataset
MSK10000
HAM10000
Class
Benign
Malignant Total
BKL
MEL
BCC
AKIEC
VASC
Total #Images Time .
In the case of the MSK10000 dataset, if we divide the overall time period of the two classes by the total number of images, we will get about 0.
044 seconds for each image.
CONCLUSIONS
This study suggested a novel Hybrid feature extraction method for the classification of skin lesions and incorporated both LBP.
GLRLM technology.
We have performed experiments using the proposed method integrated with traditional ML classifiers like SVM, kNN and Decision Tree on two publicly available datasets.
MSK10000 .
inary classificatio.
and HAM10000 .
ulti-class classificatio.
The system demonstrated strong performance on a variety of evaluation metrics, including an F1-score that reached 52% A 0.
16 for binary classification and 95.
29% A 0.
94 for multi-class tasks.
Moreover, the proposed method exhibited high computational efficiency that would allow for actual deployment in resource-limited scenarios.
One limitation of our study is that it focuses on applying the proposed hybrid feature extraction method and machine learning-based system on two specific datasets: HAM10000 and MSK10000.
While these datasets are widely used and representative of skin cancer cases, the generalizability of our approach to other datasets or real-world clinical settings needs to be further explored and validated.
Additionally, we plan to explore more sophisticated feature extraction techniques and integrate deep learning methods for further improvement.
The performance and accuracy of the proposed method may vary when applied to different datasets with variations in imaging quality, demographics, and disease prevalence.
The proposed method has a promising background as an effective and accurate system for classifying skin cancer.
However, it will need to be more extensively validated empirically with Clinical benchmarking against established evaluation frameworks.
ACKNOWLEDGEMENT
This work was supported by the Universiti Teknologi Malaysia through Geran Penyelidikan Hi-Tech(F4 ) with vote no.
: Q.
J130000.
00Q50.
REFERENCES