Lumpy Skin Disease Detection in Cattle by a Robust Approach using Advanced Convolutional Neural Networks

Cattle often suffer from skin diseases such as Lumpy Skin Disease (LSD), Dermatophilosis (Streptothricosis), Dermatophytosis (ringworm) and Bovine papillomatosis (warts) (Tageldin et al., 2014; Scott, 2018; Rojas-Anaya et al., 2016). These conditions pose a significant problem, leading to the rejection of skin and hide products due to poor quality caused by these diseases (Bekele, 2017). One of the emerging skin diseases that is posing significant health concerns for cattle worldwide is Lumpy Skin Disease (LSD).  Lumpy skin disease (LSD), which is prevalent in Africa and Turkey, has been rapidly spreading to other parts of the world, including Europe and Asia (Bekele, 2017; Whittle et al., 2023; Amin et al., 2021). This highly contagious viral disease poses a significant threat to cattle globally. LSD is caused by a virus (genus – Capripoxvirus) and has the potential to impact economic value by affecting meat and milk production, hide quality, draft power of animals and reproductive efficiency, leading to issues such as abortion and infertility (Gupta et al., 2020; Whittle et al., 2023). The disease is characterized by the formation of nodules or lumps on the skin, mucous membranes and internal organs of infected cattle. The disease may spread via a number of different routes. Through polluted settings, fomites, or vectors, there is environmental and indirect transmission. Tick cell line cultivation in vitro could help in dispersion, which might result in airborne spread. Conversely, direct contact includes the spread via contaminated medical equipment and aerosols, which may present serious dangers in medical environments. The term vertical transmission refers to the disease’s transfer from mother to child via the uterus. Additionally, vaccinated cows may sometimes discharge milk with vaccine virus, which would further complicate the dynamics of transmission. Furthermore, the disease might be expelled in the semen of cows, which increases the possibility of vertical transmission. Moreover, live animals, unsterilized animal products, or contaminated objects may transmit the disease over vast distances, which emphasizes the need for strict biosecurity measures to stop its spread.
       
A conclusive diagnosis of skin diseases requires a systematic approach that involves a complete analysis of medical history, a comprehensive physical examination and appropriate diagnostic tests. To diagnose LSD, clinicians typically rely on recognizing typical clinical signs, using differential diagnosis and employing various laboratory techniques. These techniques include serological tests, polymerase chain reaction (PCR), indirect fluorescent antibody technique and electron microscopy (Gari et al., 2008; Sharawi and Abd El-Rahim, 2011; Baselli et al., 2023). They are important, but they are not ideal for early disease screening. For example, serological assays are time-consuming and prone to potential cross-reactivity with other poxviruses (Zeedan et al., 2019). On the other hand, molecular methods such as PCR, though less time-consuming, require isolated DNA and high-end pieces of equipment and well-established setups not suitable for on-site investigation (Vidic et al., 2017). Hence, alternatives like biosensors are being developed which combine a biological component such as enzymes, antibodies, or nucleic acids with a detector to generate measurable signals (for detection) for target analytes. Specifically, PCR biosensors detect LSD viral DNA, while immunosensors use antibodies against LSD viral antigens for rapid and specific virus detection, facilitating early diagnosis and disease control in cattle populations (Li et al., 2021; Xiong et al., 2022). However, for timely and prompt measures taken to prevent the spread of disease to other animals and reduce economic loss, early detection of the disease is important. Typically, early screening involves a thorough physical examination of an animal’s hair coat and skin under strong lighting (Moriello, 2022). However, this process requires extensive human experience and expertise, which can lead to subjective errors.
       
Recently, artificial intelligence and machine learning (ML) have been integrated with the information obtained from visual inspections and biosensor data to enhance the efficiency of disease diagnosis and subsequent treatments (AlZubi, 2023; Cho, 2024; Neethirajan and Kemp, 2021; Kadian et al., 2023; Kumar et al., 2023; Wasik and Pattinson, 2024). The advent of image processing has presented an excellent opportunity for the identification of skin lesions, lumps, nodules or any other manifestation of skin diseases (Srinivasu et al., 2021; Affolter et al., 2023). In the case of LSD, the early stages of infections which are mild and may be difficult to distinguish (though the severe cases can be easily recognized) entail the utilization of artificial intelligence for disease detection with better accuracy (Tupparainen et al., 2017; Li et al., 2021; Rashid et al., 2018). The procedure is sped up with the use of image processing, which extracts distinguishing characteristics from the images (Nachbar et al., 1994). The study by Dofadar et al., (2022) used ten ML classifiers. The RF Classifier and Light Gradient Boosted Machine Classifier outperformed others, achieving a remarkable F1 score of 98%. Genemo (2023) employed CNN to develop a system trained on a Cattle’s LSD dataset by extracting features on the basis of ABCD rule of dermatoscopy which included conditions like “asymmetry (A), border (B), colour (C) and differential structure (D).
       
While existing studies demonstrate the potential of machine learning for both predicting and detecting lumpy skin disease, there remains a dearth of research utilizing image-based AI for direct lesion identification. To address this gap, this present study introduces a novel CNN model employing transfer learning for accurate lumpy skin disease detection in cattle, aiming to advance rapid diagnosis and intervention in resource-limited settings.

Image dataset
 
For the CNN model, a well-labeled, fine dataset of animal images with normal and lumpy skin disease was collected and divided into two groups representing the diseased and normal conditions. The dataset contained 1,023 images, including different categories such as age, breed and severity of disease (Table 1). Data for this study was collected with the help of specialists in the field of cattle farming. Specifically, the images in this collection included both LSD and non-LSD skin images and a sufficient number of images were kept in each category. The lumps or nodules on the skin of the animal were the main clinical indications for this disease (Fig 1). These could also be seen in the skin, cups and mucous membranes and they could range in size (Datten et al., 2023). The dataset was split into a training set and a validation set with an 80:20 ratio, where 80% of the data was used for training and 20% was used for validation.
 

 

 
Data pre-processing
 
The first step in the image-based ML technique involves the input of a dataset, known as the pre-processing stage, which helps in the optimal preparation of images by removing unwanted noise (Salvi et al., 2021). To keep images identical, they were resized to 256 × 256 pixels. This is important because the model needs uniformly sized input images to analyze data as efficiently as possible. This was followed by normalization where each image is transformed into a set of pixel values that align more closely with familiar or standard ranges (Hosakoti et al., 2021). For this, each pixel value in an image was divided by 255, the maximum value for its bit-depth, to ensure that the pixel values range between 0 and 1 (Johnson et al., 2019). Image normalization is a common procedure in image processing that alters the range of pixel intensity values. Improved integration during the model’s training is made possible by this normalization. Each picture was then labelled with the appropriate health category. These training and test datasets allowed for the identification of two classes: LSD and non-LSD skin images.
 
Data augmentation 
 
The data augmentation technique is used to generate new images based on the existing dataset to increase the diversity and variability of the dataset. Images may be shifted, rotated randomly, zoomed in and flipped vertically or horizontally during this augmentation process. The employment of data augmentation is essential in increasing the model’s capacity to generalize effectively to unknown data while simultaneously reducing overfitting problems. 
 
Model selection
 
Sequential CNN, as the pre-trained model for the classification of images, is utilized. For large dataset classification tasks, the CNN architecture has shown excellent performance. It easily identifies complex details in images.
 
Transfer learning
 
This method utilizes the trained CNN model as a feature extractor to take advantage of transfer learning. In this, the first layers’ weights are frozen and only the subsequent layers’ weights are adjusted to make them particular to the lumpy skin disease dataset. This methodology takes advantage of the depth of knowledge that the model has gained from a large dataset while conserving training time and processing resources. The procedure followed for the CNN model is presented in Fig 2.
 

 
Model architecture
 
The defined CNN model is designed for binary classification with two classes. The model architecture included six convolutional layers (Conv2D) with increasing filter sizes (32, 64) and kernel sizes (3×3), followed by max-pooling layers (MaxPooling2D) for spatial downsampling. After flattening the feature maps, the network runs data via a dense layer (Dense) activated by ReLU and 64 units. The output probabilities for each class are then generated by a dense layer using softmax activation. The architecture focuses on obtaining features through convolutional layers and records non-linear correlations in fully connected layers, following a standard pattern for image classification. Certain hyperparameters, such as layer depths and filter sizes, may require fine-tuning depending on the dataset’s properties. The important architecture and training parameters of the CNN used in this study are listed in Table 2.
 

       
Six convolutional layers build up the sequential CNN architecture and each one uses 32 or 64 filters to extract complex characteristics from the input data. Six max-pooling layers are added to the design to efficiently down-sample spatial dimensions. A regularization strategy called dropout is employed strategically at a rate of 0.5 to improve the model’s performance for generalization. The network is stable because of its uniform weight assignment, Rectified Linear Unit (ReLU) activation function and modest learning rate of 0.0001. The training protocol consists of 75 epochs with a batch size of 32 examples. Fig 3 shows the sketch of the overall architecture.
 

       
The process for the CNN model can be briefly described as follows. It involves processing the input image denoted as (x, y), with dimensions N × M, where (x, y) belongs to the set of real numbers. Equation 1 is used to compute the histogram of the image.
   
Where
h_f(k) represents the histogram of an image.
f is the frequency of events.
O_j [j = 1, 2, 3…. (k-1)] = Events of grayscales.
       
Using the above formula, the range of infected images are given using equation 2:
   
Where,
J = Pixel values.
I = Affected area.
k₁ to k_n = Range of the infected region.
~h_f(k) = Complete infected region.
       
The total image variable negatives are determined using equation 3 and 4.
   
The weight matrix and bias matrix of the convolutional layer are shown in equation 5 as follows:
   
Where
W = Weight matrix of the I^th layer.
b = Bias matrix of the I^th layer.
S = Properties of the first convolutional layer.
x, y = An enhanced image.
       
Next, the ReLu activation layer was added. The filter size of the subsequent convolutional layer was [3, 3], the stride was [1, 1] and there were 64 channels and 64 filters. ReLU activation function was used to normalize this layer’s properties. A max-pooling layer with a filter size of [2, 2] and a stride of [2, 2] was then applied. The process is compiled using six convolutional and maxpooling layers, dropout, flatten and dense layer.
 
Evaluation matrices
 
To evaluate the model’s performance, various metrics, including F1-score, Accuracy, Recall and Precision are determined. Based on the quantity of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) samples, these metrics are calculated.
 
Accuracy
 
The ratio of correctly identified tests (positive or negative) to the total number of samples is known as accuracy.
   
Precision
 
Precision, which is defined as the ratio of correct positive predictions to all positive predictions made, is a statistic used to quantify the number of correct positive predictions built.
   
Recall
 
The amount of accurate positive predictions made out of all possible positive predictions is measured using a metric called recall.
   
F1 -Score
 
A single metric called the F-score is utilized to sum up model performance. When one of the matrices is high and the other is low, it can be utilized to balance out the model’s performance.
 

In this study, a stack of convolutional layers to collect structural information and fully linked layers for classification illustrates a typical CNN approach for image classification. The architecture of the model uses max-pooling layers to gradually lower spatial dimensions, allowing the network to learn stacked representations of the input images. Class probabilities are output by the last dense layer using the softmax activation function. It’s important to note that the model’s performance is dependent on several variables, including the dataset’s characteristics, the hyperparameters selected and the particulars of the classification task. For best results, experimentation and fine-tuning with various architectures and parameters could be required. The model is trained for 50 epochs, meaning it goes through the entire training dataset 50 times during training.
       
The accuracy of training and validation datasets is shown in Fig 4(a) and the corresponding loss values are shown in Fig 4(b). After the 50 epochs, the final accuracy on the training dataset is reported as 0.8654, meaning that the model correctly classified 86.54% of events in the test dataset. The average loss over the training dataset (after 50 epochs) is 0.7323. These metrics are often utilized to measure a machine learning model’s performance on a held-out dataset and offer details on how well the model generalizes to new data.
 

       
To classify the images into actual and predicted levels, the prediction of cases belonging to a given class, such as Lumpy Skin, is presented in Fig 5. Actual Lumpy Skin, predicted lumpy skin and confidence 79.13% includes essential parts of the predictive model’s output. Actual Lumpy Skin denotes the true label of the given instance. Predicted Lumpy skin is the output produced by the model that represents the algorithm’s inference about the class identity of the instance that is being examined. The confidence score value (79.13%) represents the measured confidence level of the model in its prediction.
 

       
The metrics, including f1-score, recall and precision, are used to evaluate the CNN Model’s performance.  The confusion matrix for each class is shown in Fig 6.
 

       
The confusion matrix is a 2×2 matrix that shows the performance of a classification model performed on a binary classification test that had two classes: Normal Skin and Lumpy Skin. A classification model’s correctness can be determined by analyzing its performance using specific measures. Lumpy Skin was accurately predicted by the model nine times, as indicated by True Positives (TP). 24 incidents of the model incorrectly classifying Normal Skin as Lumpy Skin are known as False Positives (FP). False Negatives (FN) on the other hand, represent cases where Lumpy Skin has been mistakenly classified as Normal Skin (23 cases). True Negatives (TN) denote cases where the model successfully predicted Normal Skin (48 instances). These metrics together reveal the model’s ability to differentiate between Lumpy Skin and Normal Skin, with TP and TN denoting accurate predictions and FP and FN indicating regions where the model’s precision and recall might be improved.
       
The classification report (Table 3) gives a complete assessment of a model’s performance in separating between Lumpy Skin and Normal Skin classes. The precision of Normal Skin is 0.74, showing a higher precision in recognizing occurrences of Normal Skin, compared to 0.45 for Lumpy Skin, which represents the accuracy of positive predictions. The model’s recall, which measures its ability to record every relevant event, is 0.76 for Normal Skin and 0.42 for Lumpy Skin. The F1-score, which is a single statistic that combines recall and precision, is 0.75 for Normal and 0.44 for Lumpy Skin. With a weighted average F1-score of 0.65, the model’s overall accuracy is 0.65. A balanced assessment across classes is suggested by the macro-average F1-score of 0.59. However, in some classes, the model did slightly better than in others. There could be a lot of reasons for this, including the quantity of images in each class and the fundamental challenge of differentiating between them.
 

       
The current investigation, in conjunction with prior studies, demonstrates the promising outcomes attained through the application of CNN in diagnosing skin diseases. Nevertheless, a notable challenge arises when working with images captured by mobile devices or digital cameras, as CNN models are not uniformly scaled and rotation invariant. Consequently, the accuracy of diagnostic results may be compromised if the image is not captured at the appropriate angle or orientation, hindering the attainment of consistent and reliable outcomes. Subsequent research efforts are imperative to tackle this challenge and enhance the performance of CNN models in the diagnosis of skin diseases.
       
With the comparison of the presented study, Jain et al., (2022) developed an OP-DNN for human skin disease classification, achieving 95% accuracy. Mellores et al., (2020) used ML to detect dog skin diseases like yeast and ringworm. An Android app developed using OpenCV and ANN models achieved 97% and 98% prediction accuracy, respectively. Afshari Safavi (2022) used meteorological and geological factors to assess the effectiveness of machine learning algorithms in predicting LSDV infection. The Extra Trees Classifier algorithm showed remarkable accuracy of up to 97%, outperforming other machine learning methods.

The study highlights the use of image processing and machine learning techniques in the detection of animal lumpy skin disease.  A Convolutional neural network (CNN) for skin disease classification and detection is explained. The digital image processing using CNN, follows a set protocol that consists of image dataset collection, preprocessing, representation, interpretation and detection. This involves specific tasks such as ordering of channel, normalization, resizing images and data augmentation. The constructed classifier is evaluated by determining its robustness using measures like recall, accuracy, precision and F1-score. Precision and recall measures evaluate how accurately the model predicts each class in the context of this Lumpy Skin Disease study. The classification report evaluates a model’s performance in distinguishing between Lumpy Skin and Normal Skin classes. Normal Skin has a higher precision of 0.74, while Lumpy Skin has a lower accuracy of 0.45. The model’s recall and F1-score are 0.76 and 0.42, respectively, with a weighted average F1-score of 0.65. The model’s accuracy in classifying instances within the designated classes is confirmed by the combined use of all these metrics. Its potential for real-world applications is highlighted by its insightful analysis of its overall effectiveness. To improve the model’s effectiveness, experiments using a larger dataset, regularization and hyperparameter optimization, advanced image preparation methods and further experimentation on feature extraction and categorization are needed. In-depth examinations of animal skin problems are also crucial for understanding cardiovascular health.

The authors would like to thank the editors and reviewers for their review and recommendations and also to extend their thanks to King Saud University for funding this work through the Researchers Supporting Project number (RSP2024R395), King Saud University, Riyadh, Saudi Arabia.
 
Funding statement
 
This work was supported by the Researchers Supporting Project number (RSP2024R395), King Saud University, Riyadh, Saudi Arabia.
 
Author contributions
 
The author contributed toward data analysis, drafting and revising the paper and agreed to be responsible for all aspects of this work.
 
Data availability statement
 
Not applicable.
 
Declarations
 
Author(s) declare that all works are original and this manuscript has not been published in any other journal.

The author declare that they have no conflict of interest.