e-ISSN: 2798-5210

Volume 4 No. 10 Februari 2025 (3732-XXX) p-ISSN: 2798-5652

Development of Computational Models in Creating Skin Disease

Classifiers Using CNN and ResNet Architecture

Elaine Faythe Hartono1*, Putranegara Riauwindu2

University Pelita Harapan, Indonesia
*Email: Hartonoelainr@gmail.com
*Correspondence
ABSTRACT
This research paper delves into the exploration of various computational models
utilized in creating skin disease classifiers, aiming to enhance diagnostic accuracy
and optimize potential accessibility to isolated patients. By investigating the
different types of computational models employed in this context, this paper aims
to shed light on how these models contribute to more accessible and efficient
diagnoses of skin conditions, enabling healthcare professionals to provide
widespread patient care. Furthermore, the paper will address the challenges
encountered in the development of effective skin disease classifiers using
computational models, such as data quality issues, model interpretability, and
generalizability across imbalanced datasets. Through a comprehensive analysis of
these aspects, this research endeavors to advance our understanding of the
potential of computational models in improving skin disease diagnosis and
ultimately enhancing healthcare outcomes for individuals affected by
dermatological conditions. Results showed that ResNet18 demonstrated higher
overall accuracy on HAM10000 when compared to the confusion matrix and
Grad-CAM visualizations of both models.
Kata kunci: Computational models; Skin disease classifiers; CNN; ResNet;
Python.

INTRODUCTION
Skin diseases, such as acne, alopecia, bacterial skin infections, decubitus ulcers, fungal skin
diseases, pruritus, psoriasis, scabies, urticaria, viral skin diseases, and skin cancer lesions, pose a
significant public health concern worldwide, with recent years showing a rise in such diseases due to
worsening living conditions and lack of access to healthcare (Yakupu et al., 2023). Skin diseases are
the fourth most common cause of all human diseases, affecting nearly one-third of the world's
population; however, their burden is often underestimated (Flohr & Hay, 2021). Skin and subcutaneous
diseases lead to profound long-term alterations even after the disease has resolved, affecting not only
the physical health but also the mental health and quality of life of the patient, placing a high burden on
patients' families and national healthcare systems globally. The burden of skin conditions was high in
both high- and low-income countries, indicating that prevention of skin diseases should be prioritized.
There are different types of skin lesions: Actinic keratosis (AK), Basal cell carcinoma (BCC),
Benign keratosis (BKL), Dermatofibroma (DF), Melanoma (MEL), Melanocytic nevus (NV),
Squamous cell carcinoma (SCC), and Vascular lesion (VASC), are shown in Fig. 1. The lesions are
different in terms of their symptoms and severity. Some are permanent, and some are temporary and
may be painless or painful. Among these skin diseases, melanoma is the most deadly and dangerous
type. However, about 95% of skin disease patients can be recovered if identified at an initial state
(Ahammed et al., 2022). Hence, widening the knowledge of skin disease epidemiology is critical for

3732
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

policy development and resource allocation, which ultimately leads to disease prevention (Yakupu et
al., 2023).

However, in this digital age, the development of computational models has innovated the field of
dermatology by offering new approaches to skin disease classification. Through the potential concept
of integrating A.I with skin disease classification, the prevention of common skin diseases may be
accomplished. Many researchers proposed several approaches for the classification of skin diseases.
Related works are categorized into different types based on datasets, feature extraction techniques,
feature selection techniques, and classification models. This section examines several relevant research
articles to discover previous studies’ tools and procedures and identify research gaps:
Jagdish (Jagdish et al., 2022) proposed a model for skin disease detection using image processing
methods. They applied fuzzy clustering on 50 sample images with KNN and SVM classification
algorithm with wavelet analysis. They showed that the K-Nearest Neighbor classification algorithm
works well compared to the Support vector machine (SVM) classification technique with an accuracy
of 91.2% and the algorithm identified the type of skin disease using classification methods. But they
worked with only 50 sample images containing two classes (basal and squamous disease).
AlDera (AlDera & Othman, 2022) presented a skin disease diagnosis model that take an affected
skin image and diagnose acne, cherry angioma, melanoma, and psoriasis. They applied Otsu’s method
for image segmentation and Gabor, Entropy and Sobel techniques for feature extraction on the dermnet
NZ and atlas dermatologico. Finally, they applied K-Nearest Neighbor (K-NN) classifiers for
classification, and achieved 67.1% accuracy.
Janney (Janney et al., 2018) applied the Support Vector Machine (SVM), the Artificial Neural
Network (ANN), and the Naive Bayes Classifier. The features were extracted from the dermoscopic
images using the Gray Level Co-occurrence Matrix (GLCM), texture, and wavelet features. Finally, the
performance of the classification models was compared based on the accuracy, precision, and recall
values. The system achieved an average accuracy of 89%, 71%, and 71% for ANN, SVM, and Naïve
Bayes classifiers, respectively. But, the system has a poor performance on SVM and Naïve Bayes
Classifiers. It is a vital drawback of the study.
Hameed (Hameed et al., 2018) presented a novel Computer-aided Diagnosis (CAD) system for
diagnosing the most common skin lesions (acne, eczema, psoriasis, benign and malignant melanoma).
The applied Dull Razor method, Gaussian filtering for preprocessing, and Otsu’s thresholding method
for extracting the region of interest from which the features are extracted. Several colors and texture
features were extracted and applied to the Support Vector Machine (SVM) with the quadratic kernel.
The experiment was performed on 1800 images, and 83% accuracy was achieved for six-class
classification. However, the dataset was small, which is a limitation of the system.
Sinthura (Sinthura et al., 2020) proposed an advanced skin disease diagnosis model utilizing
image processing. First, they applied Otsu’s method to segment the disease portion. Next, some GLCM
features such as area, perimeter, mean, and entropy were extracted. Finally, they applied the SVM
classifier to identify the four classes of diseases: Acne, Psoriasis, Melanoma, and Rosacea. The method
identifies skin diseases with an accuracy of 89%. They applied the model to a small dataset containing
100 images and used only a small number of features. These are the drawbacks of the system.
Ubale (Ubale & Paikrao, 2019) proposed a color phase model for detecting and classifying
different skin diseases. They applied the HSV color phase model and LAB color phase model to extract

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3733
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

features from the images. Finally, they employed the K Nearest Neighbor (KNN) classification method
to classify skin diseases (Acne, Papillomas, Psoriasis, Melanoma, Mycosis, Vitiligo, Warts) that
achieved 91.80% accuracy for HSV Color Phase Model and 81.60% accuracy for LAB Color Phase
Model. Unfortunately, the preprocessing is not good enough, and there is no verified dataset used in
this study.

From the above discussions, it can be concluded that most existing research has been conducted
to detect skin diseases of a few classes. Most of the works are based on a single disease. Only a few that
have been done are inadequate for classifying multiple classes. Most of them are performed on small
size datasets. The preprocessing techniques are not good enough, and thresholding segmentation, edge-
based segmentation, clustering-based segmentation are used in most of the studies. This study attempts
to solve this limitation by examining the HAM10000 HAM10000 (“Human Against Machine with
10000 training images”) dataset, which consists of seven skin diseases with 10,015 images, and utilizes
Convolutional Neural Networks (CNN) and an SVM model-Residual Networks (ResNet) to achieve its
findings. CNNs are deep learning architectures that learn directly from data. CNNs are useful for
detecting patterns in images and recognizing objects, classes, and categories (Ahammed et al., 2022).
Residual Network (ResNet) is a deep learning model designed for computer vision applications. It is a
CNN architecture capable of supporting hundreds or thousands of convolutional layers. ResNet uses
"skip connections," which allows it to stack multiple identity mappings, skip those layers, and reuse the
activations from the previous layer, resulting in improved performance for a larger number of layers
[4].

METHOD
Data Collection and Analysis. The case data and images for skin lesions were collected from the
HAM10000 using the Harvard Dataverse online query tool. The dataset comprises of 10,015
dermatoscopic images that was publicly released by the Harvard database in June 2018 to supply
reliable training data for the automation of skin cancer lesion classification. In addition to the 10,015
images, a metadata file containing demographic information for each lesion was included, which were
verified through histopathology (histo), while the ground truth for other cases was determined through
follow-up examination (follow_up), expert consensus (consensus), or confirmation by in-vivo confocal
microscopy (confocal). (AlDera & Othman, 2022) The selection of skin lesions and diseases reflects
the prevalence, the case definition and the availability of the data used for the development of the
models.
To begin, the HAM10000 Image Dataset was first imported into the code, which contains seven
classes of skin cancer lesions: Melanocytic nevi, Melanoma, Benign keratosis-like lesions, Basal, cell
carcinoma, Actinic keratoses, Vascular lesions, Dermatofibroma. As numerical representation is
required to be usable for machine learning/artificial intelligence model, the classes must first be encoded
into categorical outcomes.

Table 1
Statistical measures after applying Gaussian filtering

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3734
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

Lesion Image dx dx_type age sex localization dataset label

HAM_0 ISIC_00336 mel histo 70.0 femal neck vidir_moder 4
004859 53 e n
HAM_0 ISIC_00281 nv histo 70.0 male upper vidir_moder 5
002856 54 extremity n
HAM_0 ISIC_00243 df consensu 40.0 femal lower vidir_molem 3
005112 86 s e extremity ax
HAM_0 ISIC_00264 nv follow_u 65.0 male lower vidir_molem 5
005431 13 p extremity ax
HAM_0 ISIC_00252 nv histo 20.0 femal abdomen vidir_moder 5
002796 61 e n
HAM_0 ISIC_00295 bkl consensu 50.0 male face vidir_moder 2
002096 18 s n
HAM_0 ISIC_00261 nv Follow_u 55.0 male trunk vidir_molem 5
002524 84 p ax

Standardization and Optimization of Dataset. To overcome the challenge of limited medical data
and improve our model's learning, a technique called data standardization and data augmentation was
used. Through PyTorch’s function that combines multiple transformations into a single callable object,
we adjusted the images' brightness, size, and colors using standard values from a large image database.
Since all our images follow a similar pattern, the model will be able to learn at a faster rate with a
simpler architecture. Next, data augmentation was applied to the training images. This involved creating
variations of our images by flipping them horizontally and rotating them, which assists our model in
becoming better at recognizing different skin conditions.
Other than augmenting and standardizing the image, the dataset was carefully organized into
distinct groups to ensure effective learning. This process involves creating three essential sets: the 'train'
set, the 'validation' set, and the 'test' set for evaluating its ability to recognize new, unseen images. By
defining a custom sampler, StratifiedSampler, used for stratified sampling of data, it takes class_vector
as input, which represents the class labels of the data. The test_size parameter determines the proportion
of data to be held out for testing. The gen_sample_array method generates stratified train-test splits
based on the class distribution. The sampler ensures that each split has approximately the same
distribution of classes as the original data. It then splits the dataset into training, validation, and test sets
using these generated indices. Finally, it prints the number of images in each set, fully rotated and
loaded into memory.

RESULTS AND DISCUSSION

The results were measured using three parameters, which were using the validation accuracy,
confusion matrix and Grad-Cam visualization. It can be concluded from the data collected that the
SVM-ResNet18 produced higher overall accuracy on HAM10000 than CNN LeNet when used to create
skin disease classifiers, making it a more ideal architecture to use when under budget or time constraints.
Accuracy and Training Loss. The results of both the CNN LeNet and ResNet 18 can be seen in
Table 3. The two models’ training and accuracy loss reveals the advantages and disadvantages of each
model. The obtained overall accuracy surpasses chance, which is approximately 14% accuracy when

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3735
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

randomly selecting a class out of the 7 available classes. This is calculated based on the expectation that
random chance accuracy can be expressed as:

Accuracy by Chance = 1 / Number of Classes

In this case, with 7 classes, accuracy by chance would be 14.29%. Both models exceeded this
and therefore fulfilled adequate benchmarks for AI learning. As the epochs increased, the respective
accuracy rates increased proportionally, with SVM-ResNet showing a faster growth when provided the
same amount of time and sample data.

Table 5
Performance metrics of our proposed models
Accuracy Rate
Models
10 Epochs 50 Epochs
CNN LeNet 33 % 54 %
SVM- 67 % 79 %
ResNet18

Confusion Matrixes. The confusion matrices of the CNN LeNet and ResNet18 models can be
seen in Figures 14 and 15. A confusion matrix is a table that summarizes the accuracy of the model's
predictions, showing the number of true positive, true negative, false positive, and false negative
predictions for each class. This gives a detailed breakdown of the model's classification accuracy and
potential areas for misclassification across classes. The confusion matrix consists of diagonal and off-
diagonal cells. The diagonal cells show the number of correct predictions in each class. The lighter or
higher the value in these cells, the more accurately the model classifies instances from that class. Off-
diagonal cells represent misclassification. Darker or lower values in these cells indicate instances where
the model made errors [14]. To generate each confusion matrix, both CNN LeNet (net) and ResNet18
(net) are applied to a test dataset (test_data_loader), and the resulting predictions are compared with
the actual labels.

Figure 1
Confusion matrix of the CNN LeNet model

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3736
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

Figure 2
Confusion matrix of the ResNet18 model

Through the confusion matrix, it’s clear that the ResNet18 model achieved more accurate results
compared to its CNN LeNet model counterpart. Figure 14 shows that the CNN LeNet model achieved
the highest performance value at identifying classes melanocytic nevi and basal cell carcinoma because
they have the highest values in their diagonal cells, as indicated by their bright color in comparison to
the rest of the classes. However, many of the off-diagonal cells are lower values, indicating that the
model made mistakes. As illustrated in Figure 15, ResNet18 outperforms the initial CNN LeNet model,
as evidenced by brighter diagonal cells. This demonstrates how the model more accurately classifies
the image provided to it based on the actual conditions. The ResNet18 model achieved the highest
performance in classifying the nv class, as evidenced by reaching the highest value. The brighter off-
diagonal cells also indicate fewer mistakes were made in comparison to the CNN LeNet model, mel
and bcc. Overall, this shows that the ResNet18 model works more efficiently with fewer errors
compared to the CNN LeNet model when classifying skin conditions.
Grad-Cam Visualizations. The confusion matrices of the CNN LeNet and ResNet18 models can
be seen in Figures 9 and 10. GradCAM is a visualization technique for interpreting the decision-making
process of CNNs, particularly in image classification tasks. It generates heatmaps, highlighting regions
of an input image that strongly influence the model's predictions. It helps highlight which features the
model focuses on during classification. GradCAM enhances the explainability and trustworthiness of
CNN-based image classification models [15].

Figure 3
Grad-CAM visualizations of the CNN LeNet model

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3737
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

Figure 4
Grad-CAM visualizations of the ResNet18 model

The images on the left in both figures are Grad-CAM visualizations created for the specified
target layer ("conv2") and target class. This visualization is created using the Grad-CAM technique on
a heat map-style visual. It highlights the regions of the original image that made the most significant
contribution to the model's prediction for the specified class. The warmer the color, the more important
that region is to the model's ability to predict the image's classification.
The image on the right in both figures is the original image of the skin diseases used. Through
comparing the ground truth label of the actual image with the prediction made in the Grad-CAM in both
the CNN LeNet and ResNet18 models, both models can correctly predict the skin condition. As seen in
both figures, the Grad-CAM can represent the shape of the skin conditions through the differences in
colors. When comparing the Grad-CAM visualizations of both models, they performed similarly, as
both models correctly predicted the skin conditions. However, ResNet18 was able to highlight the center
parts of the image in warmer colors, showing that it can identify important ‘zones’ more efficiently and
precisely, as seen in Figure 13. The CNN model on the other hand highlighted the surrounding skin,
rather than the actual lesion, demonstrating a less precise identification. Therefore, ResNet18 has shown
to have more accurate results overall on HAM10000.
Comparative Analysis. We have compared our result with some well-established recent methods
including (Janney et al., 2018), (Hameed et al., 2018), (Sinthura et al., 2020) and (Ubale & Paikrao,
2019) shown in Table 5. Our technique performed on equal levels against these state-of-the-art methods
but also required less time and costs. Complexity of a convolutional neural network model is O
(k*n*d2), where, n is the sequence length, d is the representation dimension and k is the kernel size of
convolution. On the other hand, we have used CNN and SVM. Each of these methods have less
computational complexity than deep learning-based models. Moreover, our model has shown equal
accuracy than the mentioned methods while spanning versatility.

Table 6
Comparative Analysis of the proposed model

Model Dataset Classifier Classes Accuracy

Proposed HAM10000 CNN, 7 54%,
Method SVM 79%*
Jagdish ISIC KNN, 2 91.2%
et al. SVM

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3738
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

AlDera Other CNN, 4 90.7%,

et al. Datasets SVM 84.2%,
67.1%
Janney ISIC SVM, 3 71%
et al. ANN,
Naive
Bayes
Hameed ISIC SVM 6 83%
et al.
Sinthura Unverified SVM 4 89%
et al.
Ubale Unverified KNN 7 91%
et al.

CONCLUSION
Skin disease is currently a global problem. People of many countries or regions suffer from
different types of skin diseases. We can fight against these diseases by developing various techniques
and processes. In this research, we have performed the work with several phases. We developed a skin
classifier that combines traditional classifiers such as CNN and SVM, and programming by Python to
help developers of all backgrounds potentially build upon it. We used a benchmark publicly available
dataset, which is HAM10000. As, these datasets are somehow imbalanced, so we performed data
balancing using random over sampling technique. We obtained an average accuracy of 79% and 54%
for the HAM10000 dataset using SVM and CNN classifiers, respectively. It indicates that our code
performs better using the SVM-ResNet18 model. We can use this model for other skin disease
classification tasks.
But still, there is some scope to improve the classification performance. We had limited samples
of medical data on the skin illnesses, which may lead to a narrower and less accurate scope of
identifiable diseases by the training model. As seen in Figure 7 and Figure 9, both models experienced
fluctuating rates of validation accuracy, despite the training loss decreasing. The lack of variety in the
samples provided may have led to the models learning the samples, rather than the characteristics of
identifying the illnesses. Furthermore, being able to achieve an adequate accuracy with the models
would require a substantial amount of time. Hopefully, SVM-ResNet-18 architecture could be used
more often as a learning model to assist healthcare professionals in a way that the public can always
access anywhere (i.e. an app, etc.). This way, we can gain more information on skin diseases to tackle

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3739
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

the public health concerns they cause worldwide. The code of this project is made public to encourage
further improvements.

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3740
Elaine Faythe Hartono, Putranegara Riauwindu
Development of Computational Models in Creating Skin Disease Classifiers Using CNN and ResNet
Architecture

BIBLIOGRAPHY

Ahammed, M., Al Mamun, M., & Uddin, M. S. (2022). A machine learning approach for skin disease
detection and classification using image segmentation. Healthcare Analytics, 2, 100122.

AlDera, S. A., & Othman, M. T. Ben. (2022). A model for classification and diagnosis of skin disease
using machine learning and image processing techniques. International Journal of Advanced
Computer Science and Applications, 13(5).

Flohr, C., & Hay, Rjbj. (2021). Putting the burden of skin diseases on the global map. In British Journal
of Dermatology (Vol. 184, Issue 2, pp. 189–190). Blackwell Publishing Ltd Oxford, UK.

Hameed, N., Shabut, A. M., & Hossain, M. A. (2018). Multi-class skin diseases classification using
deep convolutional neural network and support vector machine. 2018 12th International
Conference on Software, Knowledge, Information Management & Applications (SKIMA), 1–7.

Jagdish, M., Gualán Guamangate, S. P., López, M. A. G., De La Cruz-Vargas, J. A., & Camacho, M.
E. R. (2022). Advance study of skin diseases detection using image processing methods.

Janney, B. J., Roslin, S. E., & Shelcy, M. J. (2018). A comparative analysis of skin cancer detection
based on svm, ann and naive bayes classifier. 2018 International Conference on Recent
Innovations in Electrical, Electronics & Communication Engineering (ICRIEECE), 1694–1699.

Sinthura, S. S., Sharon, K. R., Bhavani, G., Mounika, L., & Joshika, B. (2020). Advanced skin diseases
diagnosis leveraging image processing. 2020 International Conference on Electronics and
Sustainable Communication Systems (ICESC), 440–444.

Ubale, A. V, & Paikrao, P. L. (2019). Detection and classification of skin diseases using different color
phase models. International Research Journal of Engineering and Technology (IRJET), 6(7).

Yakupu, A., Aimaier, R., Yuan, B., Chen, B., Cheng, J., Zhao, Y., Peng, Y., Dong, J., & Lu, S. (2023).
The burden of skin and subcutaneous diseases: findings from the global burden of disease study
2019. Frontiers in Public Health, 11, 1145513.

© 2025 by the authors. Submitted for possible open-access publication under the terms
and conditions of the Creative Commons Attribution (CC BY SA) license
(https://creativecommons.org/licenses/by-sa/4.0/).

COMSERVA: (Jurnal Penelitian dan Pengabdian Masyarakat) - Vol. 4 (10) Februari 2025- (3732-3741)

3741