Cattle can contract lumpy skin disease (LSD), a highly contagious viral illness that causes skin nodules, fever, reduced milk production and large financial losses for the livestock sector. LSD is a transboundary animal disease that has presented a significant obstacle to veterinary health infrastructure in a number of areas (Choudhari et al., 2020). Controlling the disease and reducing financial losses depend on the early identification of infected animals and their isolation.
       
Conventional diagnosis relies mostly on skilled veterinarians’ visual observations supported by laboratory testing such as PCR, ELISA and serological assays (Choudhari et al., 2020). Despite its effectiveness, this method can be time-consuming, arbitrary and less useful in rural or large-scale environments. Computer-assisted image analysis has become a feasible substitute for automated disease screening due to the growing accessibility of digital imaging technology (Saqib et al., 2024; Alam et al., 2025; Patil et al., 2025). Recent studies have also demonstrated the effectiveness of AI-driven approaches in veterinary disease diagnosis, highlighting improved accuracy and efficiency in detecting conditions such as lumpy skin disease (Alkhanifer and AlZubi, 2025).
       
Automated disease classification has greatly improved with the advent of machine learning and deep learning algorithms. Using manually created features, early research on LSD detection relied on conventional machine learning models like Random Forest and Support Vector Machines (Ujjwal et al., 2022; Kumar et al., 2023; Olaniyan et al., 2023; Patil et al., 2023). Although moderate results were obtained, these methods’ reliance on manually created descriptors made them susceptible to changes in lesion patterns, lighting and environmental complexity. According to Pal et al., (2024), the feature design continued to have a significant impact on classification accuracy.
       
A significant breakthrough in image-based LSD detection was made possible by the introduction of convolutional neural networks (CNNs). Deep CNN models have been used in a number of recent studies to distinguish between images of healthy and infected (Patil et al., 2024; Senthilkumar et al., 2024; Shakeel et al., 2024; Velugoti et al., 2024). CNN-based approaches enable automatic extraction of discriminative features from images and improve early disease detection accuracy (AlZubi, 2024). In addition, efficient deep learning architectures such as MobileNetV3 have been proposed to achieve high accuracy with reduced computational complexity, making them suitable for real-world and resource-constrained applications (Howard et al., 2019). Girmaw (2025) used EfficientNet architectures inspired by compound scaling principles (Tan and Le, 2019) to achieve high classification accuracy, while Shakeel et al., (2024) showed effective early detection using CNN models. CNN-based methods for severity-level detection were investigated in other studies, such as Narayan et al., (2023) and Kumar et al., (2024).
       
Hybrid and feature fusion approaches have also been investigated in addition to single-backbone architectures. A deep feature fusion framework that combines multiple CNN backbones, dimensionality reduction and classification was proposed by Raj et al., (2023). Similar research has been done on ensemble learning techniques to increase robustness (Ananda and Prasanna-Kumar, 2025; Khaskheli et al., 2025). The majority of these techniques relied on ensemble averaging or static concatenation, where feature contributions were not adaptively weighted, even though they occasionally improved discriminative performance.
       
Modern image classification systems have been impacted by recent architectural developments such as transformer-based feature representations (Dosovitskiy et al., 2021), compound scaling models like EfficientNet (Tan and Le, 2019) and residual learning (He et al., 2016). Lightweight and efficient model designs have also been explored to balance accuracy and computational cost, particularly for deployment in practical environments (Un et al., 2019). Interpretability and feature focus have been further enhanced by attention mechanisms and gradient-based localization methods like Grad-CAM (Selvaraju et al., 2019). However, the majority of LSD detection models currently in use do not specifically incorporate adaptive attention-driven fusion mechanisms; instead, they rely on simple feature concatenation or single CNN backbones.
       
The main limitations of existing approaches include lack of hierarchical modeling in classical methods, architectural bias in single-backbone CNNs, reliance on static fusion strategies and limited focus on reducing false negatives.
       
In order to solve these problems, this work presents a Dual-Path Attention Fusion Network (DPAF-Net) for LSD image classification. The suggested model combines complementary deep features from two convolutional backbones through channel attention and gated adaptive fusion, allowing the dynamic weighting of features and the identification of different lesions to be more accurate thus lowering misclassification.

This study was carried out at Babu Banarasi Das University, Lucknow, India. The experimental work and model development were conducted during the period July 2025 -February 2026.

Proposed methodology
 
Existing LSD detection systems based on single-backbone CNN architectures (Shakeel et al., 2024; Girmaw, 2025) have demonstrated promising performance; however, their representational capacity remains limited by architectural bias. Similarly, most feature fusion strategies rely on static concatenation or ensemble averaging (Raj et al., 2023; Khaskheli et al., 2025), without adaptively adjusting feature contributions during inference.
       
In order to overcome these shortcomings, a dual-path attention fusion network (DPAF-Net) is introduced for the automatic LSD image classification. The design includes two parallel convolutional backbones that extract complementary feature representations and combine them via channel attention and gated adaptive fusion. The entire setup is depicted in Fig 1.


 
Dual-path feature extraction
 
Let the input image be denoted as:
 

I ∈ R^{H × W × 3}

 
Where H = W = 224.
Two parallel convolutional backbones are employed:
• An EfficientNet-B4-inspired branch.
• A ConvNeXt-Tiny-inspired branch.
       
The extracted feature representations are defined as:
 

F1 = f1 (I)

 

F2 = f1 (I)

 
Where,
 

F1 ∈ R^d1

 

F2 ∈ R^d2

 
In the above equations, f1(•) and f2(•) represent the nonlinear feature extraction functions realized by the two convolutional backbones.
 
Channel attention mechanism
 
To highlight the relevant feature channels and downplay the background information, channel attention is applied separately to each feature vector.
       
For each feature vector Fi (i = 1, 2), attention weights are computed as:
 

Ai = Sigmoid [W2 • ReLU (W1 • Fi)]

 
The refined feature representation is then computed as:

 
Gated adaptive feature fusion
 
Unlike the static concatenation methods (Raj et al., 2023),  which just line up features, the new framework builds in a gating component that is learnable and works for adaptive feature fusion.
       
Initially, the features that are attended to are concatenated:
 

 
Where,
 

F_concat = ∈ R^d1 + ^d2

A gating vector is computed as:
 

G = Sigmoid (Wg · F_concat).

 
The fused representation is obtained as:
 

F_fused = F_concat    G.

 
Classification head
 
The combined feature vector goes through a fully connected classification layer:
 

Z = Wc . F_fused + b

 
For a binary classification (Healthy vs. Lumpy), the predicted probability of the positive class is obtained through the sigmoid function:
 

pi = Sigmoid (Zi)

 
Binary cross-entropy loss
 
Since the task is binary, the model is trained using Binary Cross Entropy (BCE) loss:
 

L =  - (1/N) Σ-{i = 1 to N} [yi log (pi) + (1 - yi) log (1 - pi)]

 
Where,
yi ∈ {0,1}= The ground-truth label and N is the batch size.
 
Experimental setup
 
Dataset construction
 
Because big, well, curated LSD datasets are hardly available, a consolidated dataset from multiple sources was constructed by merging publicly available Kaggle datasets with frames extracted from videos taken under field conditions and available to the public. Only videos indisputably displaying visible LSD symptoms were taken into account. A diagram illustrating the dataset construction process is presented in Fig 2.


 
Video-level splitting and frame extraction
 
To prevent data leakage, a strict video-level splitting protocol was applied. Videos were first divided into training, validation and test sets and frame extraction was performed only after this assignment. Frames were extracted at fixed rate (1 frame/second) to capture variations in lesion appearance, camera angle and lighting conditions while reducing redundancy.
 
Data cleaning and splitting
 
A multi-stage cleaning process was implemented. Images without visible cattle were removed. Exact duplicates were eliminated using hash-based filtering and near-duplicate images were filtered using similarity-based measures. No lesion-based filtering was performed in order to preserve variability.
       
The final dataset consisted of two classes: Healthy and Lumpy. The data were divided into 70% training, 15% validation and 15% testing sets. All models were evaluated on the same split.
 
Preprocessing and training configuration
 
All images were resized to 224 × 224 pixels. Data augmentation (horizontal flipping and small-angle rotation) was applied only to the training set. Validation and test sets were not augmented.
       
Models were implemented in PyTorch with ImageNet-initialized weights. Training was performed using the Adam optimizer with an initial learning rate of 0.0001, batch size of 16 and 10-12 epochs. The best-performing model on the validation set was evaluated on the test set.
 
Evaluation metrics and implementation details
 
The model’s performance was evaluated using a set of metrics; Accuracy, precision, recall, F1-score and area under the ROC curve (AUC). Since recall is a critical measure of how few negative cases are missed in disease screenings, it was given special attention.
       
Experiments were conducted in PyTorch using CUDA-enabled GPU acceleration.

Experimental design and fair evaluation protocol
 
All models were evaluated under identical preprocessing, dataset splits and training conditions. No hyperparameter tuning was performed on the test set to ensure fair comparison. 
 
Quantitative performance evaluation
 
The comparative performance of all models on the test set is presented in Table 1.


       
The overall accuracy of the proposed DPAF-Net is the largest one compared with all the models discussed (98.98%). What is more important is that it has the greatest recall (99.27%), thus the false negative rate is reduced.
       
The importance of recall is especially relevant to the livestock disease detection. The absence of a single ill animal can be the cause of infecting the disease and a false positive will only need a further test. Therefore, the increase in the recall with the proposed model not only represents a statistical advancement but also has an epidemiological meaning.
       
Despite the competitive results of the form of the Xception-based model implementation Shakeel et al., (2024) and the good recall of the EfficientNet-B7 model Girmaw (2025), the proposed model demonstrates the increase in recall with a high precision and a general trade-off.
       
The confusion matrices (Fig 3) give a class-level perspective on the prediction behavior. Compared to the baseline models, DPAF-Net has fewer false negatives and a low false positive rate. This verifies that the performance improvement is not solely due to accuracy enhancement but also due to better class-level discrimination.


 
ROC curve analysis
 
ROC curves were generated to evaluate model discrimination ability across varying thresholds.
       
As shown in Fig 4, all the models that were tested had AUC values exceeding 0. 99, which means that the models can very well distinguish between the healthy and infected classes. The model that has the highest AUC value is the proposed DPAF-Net, with a value of 0. 9992.


       
The ROC curve of DPAF-Net is near the top, left corner of the ROC space all the time.
 
Training dynamics and generalization behavior
 
To evaluate convergence behavior and optimization stability, validation accuracy and loss trends were analyzed.
       
The validation accuracy curve (Fig 5) shows smooth and stable convergence. The validation accuracy grows very quickly in the initial epochs and becomes stable after around epoch 8, which shows successful feature learning without any long-lasting instability.


       
The training and validation loss curve (Fig 6) also verifies that:
• There is a continuous decrease in the training loss.
• There is no divergence between the training and validation loss.
• There is no sign of overfitting.


 
Comparative confusion matrix analysis
 
Fig 3 illustrates the confusion matrices of all compared models under the same testing environment.
       
Based on the analysis of the confusion matrices, the following points can be noted:
• Raj et al., (2023) has a relatively higher number of false positives and false negatives.
• Shakeel et al., (2024) has better class-wise performance but still has a slightly higher misclassification rate than the previous models.
• Girmaw (2025) has a better true positive rate but has a moderate number of false positives.
• The Proposed DPAF-Net has the lowest number of false negatives and a lower number of false positives than all the other models.
In the proposed model:
• True Negatives (Healthy correctly classified): 1062.
• False Positives: 13.
• False Negatives: 5.
• True Positives: 677.
       
The very low number of false negatives directly results in the high value of recall (99.27%).
       
The comparison of the confusion matrix verifies that the performance advantage of DPAF-Net is not only in terms of accuracy but also in terms of better class-wise distribution and error reduction.
 
Failure case analysis
 
To further understand model limitations, misclassified samples were examined.

Representative misclassification examples are shown in Fig 7.


 
Observations
 
False Positives are mainly caused by:
• Irregular patches of pigmentation.
• Strong shadowing.
• Skin discoloration outside the lesion area.

False Negatives are found in:
• Early lesions with subtle texture changes.
• Distant or low-resolution images.
• Mildly visible swelling.
       
These observations indicate that the errors of the model are mainly due to the ambiguity of the images and not due to instability in the model.
 
Architectural impact on performance
 
The reasons for the enhanced performance of DPAF-Net can be summarized as follows:
Feature extraction through dual paths:
• Fine texture details are extracted by EfficientNet-B4.
• Structural and context-based features are extracted by ConvNeXt-Tiny.
Channel attention:
• Boosts the channels of relevant features.
Gated feature fusion:
• Complementary feature maps are dynamically fused.
• Redundancy is minimized, unlike simple concatenation.
       
The organized fusion strategy allows for more comprehensive multi-scale representation learning, which is presumably responsible for the enhanced recall and AUC values.
 
Practical and clinical implications
 
The high recall (99.27%) indicates improved sensitivity for detecting infected cattle, while maintaining high precision. This balance is important for practical livestock disease screening systems.
 
Summary of findings
 
The experimental outcomes show that:
• Deep CNN models perform better than traditional feature-based approaches.
• Very deep models enhance recall but can cause increased computational complexity.
• The proposed DPAF-Net model performs well.
• Attention-guided dual-path fusion improves discriminative power and suppresses false negatives.
       
In conclusion, the results validate the efficiency of adaptive dual-path feature fusion for LSD detection.

This study proposed a Dual-Path Attention Fusion Network (DPAF-Net) for automated detection of Lumpy Skin Disease (LSD) in cattle using image data. By integrating dual-path feature extraction with attention-guided gated fusion, the model effectively captures complementary deep features and improves lesion representation. Experimental results demonstrate strong performance across multiple metrics, particularly achieving high recall, which is crucial for early disease detection. The proposed approach maintains end-to-end trainability while reducing false negatives and enhancing classification reliability. These findings highlight the potential of attention-based fusion models in developing robust and scalable livestock disease detection systems. Future work can focus on extending the model to multi-class scenarios, incorporating transformer-based architectures and improving real-world applicability through diverse datasets and explainable AI techniques.

The authors declare that there is no conflict of interest regarding the publication of this paper.