© JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

CNN-Based Classification for Lumpy Skin Disease

Detection (Bovine-HealthGuard)
TOYIN OKEBULE
Department of Computing, Afe Babalola University, Ado-Ekiti, Nigeria

Abstract- Lumpy Skin Disease (LSD) is a disease that I. INTRODUCTION

affects cattle all over the world, characterized by
distinctive skin nodules caused by the nettling virus. Lumpy Skin Disease (LSD) is a significant viral
This disease poses significant financial challenges disease affecting cattle, characterized by distinct skin
due to its negative effects on milk production, skin nodules caused by the Nettling virus, it is significant
quality, and overall cattle health, Although LSD contributor to of financial losses in the livestock
originated in sub-Saharan Africa, recent outbreaks business of some countries. Lumpy Skin Disease
in Europe and the Middle East. The disease (LSD) is a terrible viral disease that mostly affects
continues to have an incredible ability to spread cattle and is identified by prominent skin nodules
across cattle worldwide. In order to prevent animal brought on by the Nettling virus. Lumpy skin disease
disease outbreaks, technological solutions are affects cow populations worldwide, particularly in
urgently required, as the present conventional Nigeria, and presents significant economic hurdles in
methods for LSD detection are time-consuming and addition to health issues. LSD outbreaks might have a
require skilled experts. By examining into the catastrophic impact on local economies and food
epidemiology, patterns of transmission, and security in Northern Nigeria, where raising cattle is a
economic effects of LSD, this study seeks to improve major source of income. At first, Lumpy Skin Disease
on detection of lumpy skin disease. With the was mistaken for poisoning or bug bite sensitivity [1].
implementation of the trending technology - Deep The possibility of additional regional extension, which
Learning, (Convolutional Neural Networks (CNNs)) might involve Nigeria's neighbors and possibly the
on lumpy skin disease image datasets for detection. whole West African region, is made clear by recent
Image preprocessing were utilized and resized to LSD breakouts in a number of locations [2]. Concerns
640x640 pixels on YOLOv5 model and 200x200 over the disease's potential effects on cattle
pixels on Xception and ResNet models. Regions of populations in Nigeria's north, where livestock
interest were identified as an adaptive thresholding husbandry is common, are raised by its appearance in
segmentation, while noise was reduced and image new areas. Blood-feeding insects, especially
intensity was balanced using Gaussian filtering and mosquitoes and stable flies, which are prevalent in
histogram equalization. Experimental results Nigeria's grasslands and savannas, are the main cause
demonstrated that YOLOv5 outperformed the two of LSD transmission. Significant economic
other models with 82.89% of accuracy compared to ramifications result from the disease, which eventually
ResNet and Xception. The study shows that the affects the lives of cattle producers in Northern
importance of web integration to enable the detection Nigeria by causing permanent damage to hides,
of widespread lumpy skin diseases in a real time impaired producing milk, loss of weight, sterility in
application. bulls, and premature delivery in pregnant cows [3].
Additionally, LSD interferes with the trade in cattle
Indexed Terms- Lumpy Skin, Convolutional Neural and their products, which hurts the economics of areas
Network, Image Processing, Bovine-Health guard, where epidemics occur. Developing successful control
Disease Detection. and preventive initiatives in Nigeria requires an
understanding of the epidemiology and mechanisms of
LSD transmission. The disease's unique clinical
symptoms require laboratory confirmation using
diagnostic techniques like polymerase chain reaction

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 966

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

(PCR) and serological testing. Researchers from all thorough analysis of earlier research was necessary for
around the world are examining the possibilities of this. The following literature review highlights
deep learning, a subset of artificial intelligence, in the important findings and contributions from previous
medical industry due to its rapid development and studies:
strong interest in applying it to medical imaging issues
[4]. Deep learning techniques, including Abunadi et al. in [8] developed deep learning and
Convolutional Neural Networks (CNNs), have shown machine learning techniques for diagnosing skin
remarkable success in image classification tasks [5]. diseases early using dermoscopy images. using ANN,
These methods can greatly increase the accuracy of FFNN and convolutional neural networks (ResNet-50,
image-based classification as they have great potential AlexNet). Pre-processing techniques were employed
to improve the early detection of medical conditions. to enhance datasets, remove noise, and eliminate hair.
Conventional detection methods usually depend on Experimental results showed the effectiveness of the
clinical symptoms, and Lumpy Skin Disease shares systems, with the ANN algorithm achieving 95.3%
clinical similarities with other cattle diseases, making and 97% accuracy on ISIC 2018 and PH2 datasets,
it difficult to distinguish based only on noticeable respectively, outperforming FFNN in the former but
symptoms [6]. The proposed diagnostic approach for not the latter. ResNet-50 and AlexNet CNN models
Cattle's Lumpy Skin Disease (CLSD) leverages deep exhibited superior performance, with ResNet-50
learning algorithms, including a 10-layer achieving accuracies of 90% and 95.8% on the ISIC
Convolutional Neural Network and color histogram- 2018 and PH2 datasets, respectively. However, the
based segmentation, coupled with an Extreme study encountered limitations related to significant
Learning Machine classifier, achieving a high feature similarity among some skin diseases,
accuracy of 90.12% [7]. This AI-based system offers potentially causing confusion for classification
a time-efficient and scalable solution, contributing to algorithms during diagnosis.
early disease detection and presenting a competitive
edge compared to traditional diagnostic methods. Another author in [9] adapted Google's EfficientNet-
Through extensive study, this study aims to build upon b4 as the CNN backbone, pre-trained on ImageNet,
the work cited in [2] and further demonstrate the and added seven auxiliary classifiers to intermediate
complexities of Lumpy Skin Disease by providing a layer groups. Retraining the modified model using
valuable insights that could guide focused on control PyTorch, the framework achieved an overall accuracy
measures. This study covers the developments in of 0.948, sensitivity of 0.934, and specificity of 0.950.
disease awareness, detection techniques, integration of Comparative analyses with existing CNN models
technology, and the broader application of artificial demonstrated superior performance, with the highest
intelligence in veterinary healthcare. The aim of this area under curve (AUC) of 0.985. However, the study
research is to add to the existing understanding on potentially reduced accuracy on unseen data, real-
Lumpy Skin Disease with a focus on its incidence, world diagnostic considerations beyond images,
transmission patterns, and economic implications. limited disease range in the dataset, and a lack of
With user-friendly web tool for the early diagnosis of exploration into web integration for broader clinical
Lumpy Skin Disease in cattle, this study additionally applications.
seeks to empower farmers by reducing the financial
effect, facilitating prompt interventions, and Muhaba et al. in [10] developed an automatic skin
improving livestock health monitoring in general. disease diagnosis system using deep learning from
Furthermore, this will benefit cow numbers around the clinical images and patient information, where clinical
world and the farming communities that depend on images were captured using various smartphone
them. cameras and patient data were collected during
registration. The study achieved a multiclass
II. LITERATURE REVIEW
classification accuracy of 97.5%, sensitivity of 97.7%,
To gain a deeper understanding and insight into the and precision of 97.7% for five common skin diseases,
most successful technique for this research, a thorough demonstrating excellent diagnostic performance.
examination of pertinent literature was conducted. A Further validation in real-world clinical settings is

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 967

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

necessary to evaluate the robustness and reliability of model generalization. Suparyati et al. [14] proposed
the proposed smartphone application. Another in [12] different sampling and oversampling with SMOTE, to
proposed a brain tumor detection system based on predict Lumpy Skin Disease (LSD) using a Random
deep learning approaches and magnetic resonance Forest algorithm. The study addressed class imbalance
imaging (MRI), utilizing an optimized YOLOv7 in the LSD dataset and standard metrics - precision,
architecture with advanced features like SPPF+, recall, F1-Score, and ROC AUC was evaluated. The
BiFPN, Decoupled Head, and Attention Mechanism. under sampling achieved balanced class distribution
The model was fine-tuned and trained on an MRI but encountered challenges with closely mixed data.
dataset with transfer learning from the COCO dataset, More so, oversampling with SMOTE significantly
achieving enhanced feature extraction and improved model performance, achieving high recall,
localization. The result demonstrated superior precision, and F1-score while reducing the false-
performance in detecting of pituitary, meningioma, negative rate. The study effectively demonstrated the
and glioma tumors with a prediction accuracy of utility of data balancing techniques for LSD prediction
99.5%. However, further validation on larger and with SMOTE showing superior performance.
more diverse datasets are required so as to ensure Bousabarah et al. in [15] proposed an automated
robust generalization on models’ performance. detection and delineation system for hepatocellular
Krishan et al. [12] proposed an ensemble model on carcinoma (HCC) on multiphasic contrast-enhanced
liver cancer detection and classification with CT MRI using a deep convolutional neural network
images. The study datasets were sourced from (DCNN) with a U-Net architecture. The study, a
hospitals and ethical clearance from radiologists retrospective single-center analysis, implemented on a
which included normal and abnormal liver images, dataset from 2010 to 2018, the datasets was partition
with features extracted from manually selected regions to training, the U-Net on 70% and validating and
of interest called (ROIs) and implemented on Adaptive testing were implemented on the remaining 30% of
Boosting, Random Forest, Support Vector Machine, datasets. With a dice similarity coefficient (DSC)
Generalized Linear Model, and Neural Network and requirement of larger than 0.2 between individual
an ensemble of these classifiers achieved higher lesions and their segmentations, the DCNN obtained
accuracy of 98.39. However, user friendly application detection rates of 73% on validation sets and 70% on
was not implemented in this study. Hussain et al. [13] testing sets, respectively. The average false positive
experimented a Computer Vision Approach for Liver rate (AFPR) was enhanced by post-processing using a
Tumor Classification Using CT Dataset, proposing a random forest (RF) classifier and thresholding (TR).
Multi-Class Liver Tumor Identification (MLTI) However, because the study depended on particular
framework for early-stage liver tumor detection. The MRI sequences, validation on larger and more varied
methodology involved dataset preprocessing, multi- datasets is necessary to assess the model's
feature extraction (histogram, texture, binary, RST), generalizability. Bassel et [Link] [16] developed an
feature optimization using correlation-based feature automatic skin cancer classification system using a
selection, and classification using machine learning hybrid deep learning approach, leveraging a dataset
algorithms (J48, RF, RT, LMT). The dataset consisted from the ISIC archive with 70% for training and 30%
of liver tumor CT images with four classes for testing, comprising 1800 benign and 1497
(hemangioma, cyst, hepatocellular carcinoma, and malignant skin cancer images. The pre-trained models
metastasis). The MLTI framework demonstrated such: ResNet50, VGG16, and Xception were used for
promising results in liver tumor classification, feature extraction, and a stacked CV model that
achieving 97.48% accuracy using optimized feature combined SVM, KNN, random forest, decision tree,
sets obtained through correlation-based feature and regression was used for classification.
selection and 10-fold cross-validation. ML classifiers Performance measures - accuracy, sensitivity, F1
- (J48, LMT, RF, and RT) were employed on regions score, and AUC score were evaluated. The result
of interest (ROIs) in liver CT images, with RF shows that Xception outperformed other models with
achieving 97.48% of accuracy on 17x17 ROIs. 90.9% of accuracy. However, this model relied on a
However, the study gave further research on single dataset and was not implemented on user-
implementing the datasets on other ML classifiers for friendly application.

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 968

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

The review addresses numerous strategies applied in

skin disease identification via image processing and
machine learning techniques. Several researchers have
built systems leveraging multiple methodologies,
including deep learning, CNN architectures, and
machine learning algorithms to detect and categorize
skin disorders. While many studies demonstrate
promising results, several common limitations
emerge, such as reliance on specific datasets, limited
exploration of computational constraints, potential
dataset dependencies, and the need for further
validation in real-world deployment scenarios.
Additionally, the studies often lack comprehensive
benchmarking against state-of-the-art methods and
may not fully address issues related to model Fig. 1: Block Diagram for the Proposed System
generalization and scalability. Overcoming these
limitations is crucial to ensuring the effectiveness and Image acquisition
reliability of skin disease identification systems in This study procured the datasets from Kaggle,
clinical practice. leveraging internet connectivity to access and
download the dataset in the PNG (Portable Network
III. MATERIALS AND METHODS Graphics) file format. To augment the dataset and
This session provided the system structured to enrich the study’s analysis, the data was expanded by
delineate the architecture, methodologies, and acquiring additional datasets from publicly available
algorithms adopted in accordance with the proposed sources using the Google search engine. Subsequently,
research framework. Initially, the study outline the the collected image data from both Kaggle and
proposed system architecture in Fig. 1, subsequent supplementary sources, focused on extracting relevant
sections delve into the explanation of methodologies information related to Lumpy Skin Disease. The
and algorithms utilized for executing and evaluating review addresses numerous strategies applied in skin
various stages of the process, including image disease identification via image processing and
preprocessing, image segmentation, feature machine learning techniques. Several researchers have
extraction, and classification. The primary objective of built systems leveraging multiple methodologies,
this study is to develop and deploy a sophisticated including deep learning, CNN architectures, and
classification system for the detection of Lumpy Skin machine learning algorithms to detect and categorize
Disease in animals, employing deep learning skin disorders.
algorithms for enhanced accuracy and efficiency.
Image Pre-Processing
The sourced image datasets goes through image
A. Dataset Preparation
preprocessing, this is a crucial stage that improves the
The Data set consists of a total of 300 lumpy skin and
original image dataset’s quality by eliminating
700 normal Cow images from Kaggle and through
unnecessary elements like background noise. The
Google search. These images were captured using a
following image dataset preprocessing were carried
digital mobile camera under specific conditions and
out in this study:
this undergo several processes as discussed below:

Resizing the image

Each image were first downsized to a standard of 640
by 640 pixel size so as to maintain consistency
throughout the dataset, this scaling made compatibility
and reliable feature extraction possible. There reasons
for this resizing are: first, it made the image more

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 969

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

compatible so that they could be easily incorporated D. Model redeployment

into the analysis process. Secondly, the amount of i. ResNet Model
characteristics retrieved from each image was unified The ResNet model was implemented on the image
by standardizing the image size, which made the next datasets through a convolutional layer which contains
processing stages easier for this study. 64 filters, kernel size of 7x7, and stride 2, the batch
normalization and ReLU were also carried out. The
.Auto-adjusting contrast output was then down sampled using max pooling
Machine learning models are dynamically modify the with a kernel size of 3x3 and stride 2. The output
different input data while training are known as auto- which then passed through a series of residual blocks,
adjusting the difference training models. Through this that consist of two convolutional layers, batch
technique, an image's appearance can be changed normalization, and ReLU, on the other hand, the
without affecting its content by adjusting the intensity residual connection add up the input together with the
distribution of its pixel values. By altering the contrast, output. By this, the model to learn residual functions
the models were able to identify objects in a variety of that ease the training process and then improved the
lighting scenarios, backdrops, and other performance of the model. The output passes through
environmental aspects, such as generalization and a final convolutional layer with a kernel size of 1x1
increased model robustness. and a softmax activation function for classification.
Resent is calculated as the following Eqs. 1 to 7.
Auto orient
For consistent processing of images taken from The Input image is given as Eq. 1 while output Y of a
various positions, automatic orientation is a feature convolutional layer is given as Eq. 2.
that was added to the image datasets so that 𝑿𝝐𝑹𝑯×𝑾×𝑪 (1)
it automatically aligns images to a desired alignment. 𝑭−𝟏 𝑭−𝟏 𝒄−𝟏
𝒀𝒊,𝒋,𝒌 = ∑𝒎=𝟎 ∑𝒏=𝟎 ∑𝒄=𝟎 𝑾𝒎.𝒏.𝒄,𝒌 (2)
Orientation correction, text recognition, image Where:
annotation and processing pipelines, algorithmic 𝑭 is the filter size (7), 𝑺 is the stride (2), 𝑾 are the
approaches, and integration with image processing weights of the filters, 𝑪 is the number of input
libraries are some of the aspects that fall under this channels,
category. 𝑲 indexes the output filters (64) ans 𝑯 is Height of the
input.
B. Data Splitting The Batch normalization normalizes the activations of
In order to ensure the reliability and consistency of the the previous layer for independent mini-batch. Given
convolutional neural network (CNN) models for the an activation 𝒙 over a mini-batch 𝐁 = {𝒙𝟏. . . 𝒎}: in
detection of lumpy skin disease classification, a data Eq. 3 to 7.
splitting approach was used. Three subsets of the 𝟏
µ𝛃 = ∑𝒎
𝒊=𝟎 𝒙𝒊 (3)
dataset were created - 70% of the data was used for 𝒎
𝟏
training, 20% was used for validation, and the 𝟐
𝝈 = ∑𝒎 (𝒙𝒊 − µ𝛃) 𝟐
(4)
𝒎 𝒊=𝟎
remaining 10% was used for testing. â𝟐 =𝒙𝒊 −µ𝛃
(5)
√𝝈𝟐
𝛃 +𝝐
C. Training and Testing Phase
𝒚𝒊 = 𝜸â𝟐 − 𝛃 (6)
In the training phase, three different CNN models were
employed: YOLOv5, Inception, and ResNet. Each of
Where 𝜸 and 𝛃 are learned parameters, and 𝝐 is a
these models implemented on the training data subset
small constant for numerical stability.
to learn the underlying patterns and features associated
with lumpy skin disease lesions. After the training, the
𝒇(𝒙) = 𝐦𝐚𝐱(𝟎, 𝒙) (7)
performance of the trained models was evaluated
In case the dimensions of 𝑭(𝒙) = and 𝒙 do not match,
using the testing subset. The evaluation metrics
a linear projection 𝑾𝒙 can replace this and this is
employed includes the Accuracy, F1-Score, Precision,
given as Eq. 8:
and Recall.

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 970

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

be efficient and scalable, making it suitable for a wide

𝒚 = 𝑭(𝒙) + 𝑾𝒔𝒙 𝒙 (8) range of image classification tasks. The Xception
The softmax function is used for classification and is model uses depthwise and Pointwise convolutions
given b Eq. 9 which is given as Eq. 11 and 12:
𝒆𝟐 𝒋
𝝈(𝒁)𝒊 = ∑𝒌 (9)
𝒌=𝟎 𝒆𝟐 𝒌
𝑦𝑑𝑤 = 𝑊𝑑𝑤 ∗ 𝑥 (11)
𝑦𝑝𝑤 = 𝑊𝑝𝑤 ∗ 𝑦𝑑𝑤 (12)
Where 𝒛 is the input to the softmax is function, and 𝒌
is the number of classes. Where:
𝑦𝑑𝑤 𝑎𝑛𝑑 𝑦𝑝𝑤 are the output feature maps of the
ii. YOLOv5 Model depthwise and pointwise convolutions, respectively.
In this study the YOLOv5 model were implemented 𝑊𝑑𝑤 iand 𝑊𝑝𝑤 represent the weights (filters) for
on the input image datasets by passing the input image depthwise and pointwise convolutions, respectively.
through a series of convolutional and CSP blocks to 𝑥 represent input feature map
extract the features contained. The output then passes ∗ represent the Convolution operation
through a prediction layer which predicts the bounding
box coordinates, confidence score, and class D. Model Evaluation
probabilities. The predicted bounding boxes are then The model's performance was assessed using accuracy
filtered using non-maximum suppression to remove and loss metrics during both training and testing
redundant boxes. The model uses a loss function that phases. Confusion matrices were employed to
combines these bounding box loss, confidence loss, examine individual class predictions. Furthermore,
and classification loss in order to optimize the model's distinct comparisons were made between the Xception
performance which allowed the model to detect the and ResNet models regarding their accuracy during
objects in real-time with high accuracy. ResNet model training and testing. To assess the performance of the
is a deep neural network that uses residual connections stated models in the study, employed several
to reduce complexity during the training process as evaluation metrics were employed including,
this given by Eq. 10: Accuracy, Precision, Recall, and F-score indicated in
Eqs. 13 to 16.
𝐹(𝑥) = 𝑊2 ∗ ReLU(BN(𝑊1 ∗ 𝑥 + 𝑏1 )) + 𝑏2 (10)
Where: 𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (13)
𝑥 represents the input feature map, 𝑊1 and 𝑊2 are 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

the weights of the filters, 𝒃 represents the biases, 𝑇𝑃

𝐑𝐞𝐋𝐔 represents the Rectified Linear Unit activation 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (14)
𝑇𝑃+𝐹𝑃
function, BN represents Batch normalization function,
𝑇𝑃
𝑏1 and 𝑏2 are biases for the two convolutional layers. 𝑅𝑒𝑐𝑎𝑙𝑙 = (15)
𝑇𝑃+𝐹𝑁

iii. Xception Model 2∗𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙

𝐹1 − 𝑠𝑐𝑜𝑟𝑒 = (16)
The Xception model was implemented by passing the 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙

input image through a series of convolutional and max Where:

pooling layers so as to extract the features contained. 𝑇𝑃 means True Positive, 𝑇𝑁 means True Negative, 𝐹𝑃
The output derived from this was passes through a means False Positive and 𝐹𝑁 means False Negative.
series of depthwise separable convolutional layers, The result of standard metrics evaluations for the three
that consist of a depthwise convolution with the deep leaning learning models employed in this study
pointwise convolution. This method made the model are depicted in Table I.
more efficient by ensuring that it could extract features
with less data. Furthermore, a softmax activation E. Web Application
function was applied to the output in order to classify In addition to the development of convolutional neural
it accurately. The model's architecture is designed to network (CNN) models for lumpy skin disease
classification, this study encompasses the integration

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 971

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

of these models into a user-friendly web application. G. System Deployment

The web integration serves as a practical platform for The study was implemented using Visual Studio Code
users to upload images, obtain real-time classification (VS Code), an Integrated Development Environment
results for lumpy skin disease lesions, and access (IDE) that is suitable for deep learning tasks in this
educational resources related to lumpy skin disease. study. The web application was built on the Django
framework, so as to provide a robust features for
F. Educational Resources efficient web development and deployment of deep
In addition to facilitating image classification, the web learning models. Additionally, a database solution
application serves as a valuable educational resource such as Dynamo DB was incorporated into the system
hub for lumpy skin disease. Users can access a curated architecture to store images post-classification,
collection of educational materials, including articles, ensuring data persistence and accessibility. This
videos, and interactive tutorials, aimed at increasing comprehensive integration of technologies facilitated
awareness, understanding, and management of lumpy the seamless deployment of the deep learning model
skin disease. within a web environment, enhancing its usability and
accessibility for end-users as shown in Fig. 2 to 4.
IV. RESULT AND DISCUSSION

The implementation of this study was carried out

within a Python environment (version 3.8), utilizing
Keras with TensorFlow backend to develop and train
a CNN model. Additionally, image augmentation and
model building were facilitated using the Roboflow
platform, and PyTorch was employed for
computational tasks. The model performance analysis
in graph stated in Table I.

TABLE I. PERFORMANE ANALYSIS OF

THHE THREE MODELS EMPLOYED Figure 2: Home page
Precisio Recal F1
Model Fig. 2 depicts a user-friendly interface that allows
n l Score
users to upload images for classification
ResNet 0.643 0.703 0.671

Xception 0.705 0.758 0.731

YOLOv5 0.829 0.950 0.885

The ResNet model achieved a precision of 0.643,

recall of 0.703, and F1 score of 0.671, while the
Xception model demonstrated precision of 0.705,
recall of 0.758, and F1 score of 0.731. The YOLOv5
model outperformed the performance metrics with a
precision of 0.829, recall of 0.950, and F1 score of Figure 3: Lumpy skin classification to yes
0.885. More so, all models performed excellently, the
YOLOv5 model performance better compared to Fig.3 is the lumpy Skin Classification that accurately
ResNet and Xception in terms of precision, recall, and classifies the images as either "Yes", that is, lumpy
overall F1 score and validation accuracy. skin is detected.

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 972

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

AUTHOR’S CONTRIBUTIONS

The author conceived and designed the study,

developed the deep learning approach using
Convolutional Neural Networks (CNNs), conducted
the experiments, analyzed the results, wrote the
manuscript and approved the final submission.

ACKNOWLEDGMENT

Figure 4: Lumpy Skin classification to No

The author acknowledges with profound gratitude the
Fig. 4. Is the Lumpy Skin Classification that support and facilities provided by Afe Babalola
accurately classifies the images into category of "No" University, Nigeria, which facilitated the successful
which means lumpy skin is detected. completion of this research work.

CONCLUSION
REFERENCES
This study successfully achieved its primary aim of
[1] O. M. Olaniyan, O. J. Adetunji, and A. M.
developing an advanced Lumpy Skin Disease
Fasanya, “Development of a Model for the
detection and classification system. By leveraging
Prediction of Lumpy Skin Diseases using
cutting-edge technologies such as TensorFlow,
Machine Learning Techniques”, AJERD, vol. 6,
Django, and Python, we designed a highly effective
no. 2, pp. 100–112, 2023.
and efficient system that integrates a comprehensive
image dataset and a robust image classification model. [2] A. Mazloum, A. S. Van, S. Babiuk, E. Venter, D.
The YOLOv5 model demonstrated superior B. Wallace, & A. Sprygin “Lumpy skin disease:
performance with an accuracy reflected in its precision history, current understanding and research gaps
of 0.829, recall of 0.950, and F1 score of 0.885, in the context of recent geographic expansion”,
showcasing its potential to enhance lumpy skin disease Frontiers in Microbiology, 14, 1266759., 2023.
diagnosis. The web-based platform ensures user- [Link]
friendly accessibility, making it a valuable tool for [3] Kumar, B. Kumar and H. S. Negi, "Predicting
practical applications. Moreover, this study Lumpy Skin Disease using Various Machine
contributes significantly to the field of lumpy skin Learning Models," 2023 International
disease detection, paving way for further research Conference on Computational Intelligence and
works. Sustainable Engineering Solutions (CISES),
Greater Noida, India, 2023, pp. 412-416, doi:
RECOMMENDATION AND FUTURE WORK 10.1109/CISES58720.2023.10183604.
[4] S. Afshari & Ehsanallah. (2022). Assessing
To further improve the performance and reliability of machine learning techniques in forecasting
the lumpy skin disease detection model, this study lumpy skin disease occurrence based on
suggests expanding the dataset size to include a more meteorological and geospatial features. Trop.
comprehensive representation of diverse cases, Anim. Health Prod., 1(54), 12- 37, doi:
thereby enhancing accuracy and robustness. 10.1007/s11250-022-03073-2.
Furthermore, diversifying data sources beyond a
[5] R. Geetha, T. K. Rao, & M. E. Palanivel. (2024).
single origin would improve the model's
Machine learning based diagnosis of lumpy skin
generalizability and performance across various
disease. Journal of Emerging Technologies and
populations and environments, ultimately leading to a
Innovative Research, 11(7), 322-336.
more effective disease detection system.
[6] A. K. Jain, R. G. Tiwari, N. Ujjwal and A. Singh,
"An Explainable Machine Learning Model For

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 973

 © JUN 2025 | IRE Journals | Volume 8 Issue 12 | ISSN: 2456-8880

Lumpy Skin Disease Occurrence Detection," Skin Disease," Jurnal RESTI (Rekayasa Sistem
2022 International Conference on Data Analytics dan Teknologi Informasi), vol. 6, no. 4, pp. 555-
for Business and Industry (ICDABI), 2022, pp. 562, 2022, doi: 10.29207/resti.v6i4.4147.
503-508, 2022. doi: [15] K. Bousabarah, B. Letzen, J. Tefera, L. Savic, I.
10.1109/ICDABI56818.2022.10041665. Schobert, T. Schlachter, L. H. Staib, M. Kocher,
[7] M. Genemo, “Detecting high-risk area for lumpy J. Chapiro, and M. Lin, "Automated detection
skin disease in cattle using deep learning and delineation of hepatocellular carcinoma on
feature”, Advances in Artificial Intelligence multiphasic contrast-enhanced MRI using deep
Research, 3(1), 27–35, 2023. learning," Abdominal Radiology, 2020, doi:
[Link] 10.1007/s00261-020-02604-5.
[8] I. Abunadi and E. M. Senan, “Deep learning and [16] A. Bassel, A. B. Abdulkareem, Z. A. A.
machine learning techniques of diagnosis Alyasseri, N. S. Sani, and H. J. Mohammed,
dermoscopy images for early detection of skin "Automatic Malignant and Benign Skin Cancer
diseases,” Electronics, vol. 10, no. 24, p. 3158, Classification Using a Hybrid Deep Learning
2021. Approach," in Diagnostics, vol. 12, no. 10, pp.
[9] Zhu, C. Y., Wang, Y. K., Chen, H. P., Gao, K. 2472, 2022, doi: 10.3390/diagnostics12102472.
L., Shu, C., Wang, J. C., et al., “A deep learning
based framework for diagnosing multiple skin
diseases in a clinical environment”, Frontiers in
Medicine (Lausanne), 8, 626369, 2021.
[Link]
[10] K. A. Muhaba, K. Dese, T. M. Aga, F. T. Zewdu
and G. L. Simegn, "Automatic skin disease
diagnosis using deep learning from clinical
image and patient information," in Skin Health
and Disease, vol. 2, no. 1, pp. e81, 2021, doi:
10.1002/ski2.81.
[11] A. B. Abdusalomov, M. Mukhiddinov and T. K.
Whangbo, "Brain Tumor Detection Based on
Deep Learning Approaches and Magnetic
Resonance Imaging," in Cancers, vol. 15, no. 16,
pp. 4172, 2023, doi: 10.3390/cancers15164172.
[12] A. Krishan and D. Mittal, "Ensembled liver
cancer detection and classification using CT
images," in Proceedings of the Institution of
Mechanical Engineers, Part H: Journal of
Engineering in Medicine, vol. 235, no. 2, pp.
232-244, 2021, doi:
10.1177/0954411920971888.
[13] M. Hussain, N. Saher, and S. Qadri, "Computer
Vision Approach for Liver Tumor Classification
Using CT Dataset," in Applied Artificial
Intelligence, vol. 36, no. 1, pp. 1-23, 2022, doi:
10.1080/08839514.2022.2055395.
[14] S. Suparyati, E. Utami, and A. H. Muhammad,
"Applying Different Resampling Strategies In
Random Forest Algorithm To Predict Lumpy

IRE 1709140 ICONIC RESEARCH AND ENGINEERING JOURNALS 974