SkinVisualNet: A Hybrid Deep Learning Approach Leveraging Explainable Models for Identifying Lyme Disease from Skin Rash Images

Abstract

Lyme disease, caused by the Borrelia burgdorferi bacterium and transmitted through black-legged (deer) tick bites, is becoming increasingly prevalent globally. According to data from the Lyme Disease Association, the number of cases has surged by more than 357% over the past 15 years. According to the Infectious Disease Society of America, traditional diagnostic methods are often slow, potentially allowing bacterial proliferation and complicating early management. This study proposes a novel hybrid deep learning framework to classify Lyme disease rashes, addressing the global prevalence of the disease caused by the Borrelia burgdorferi bacterium, which is transmitted through black-legged (deer) tick bites. This study presents a novel hybrid deep learning framework for classifying Lyme disease rashes, utilizing pre-trained models (ResNet50 V2, VGG19, DenseNet201) for initial classification. By combining VGG19 and DenseNet201 architectures, we developed a hybrid model, SkinVisualNet, which achieved an impressive accuracy of 98.83%, precision of 98.45%, recall of 99.09%, and an F1 score of 98.76%. To ensure the robustness and generalizability of the model, 5-fold cross-validation (CV) was performed, generating an average validation accuracy between 98.20% and 98.92%. Incorporating image preprocessing techniques such as gamma correction, contrast stretching and data augmentation led to a 10–13% improvement in model accuracy, significantly enhancing its ability to generalize across various conditions and improving overall performance. To improve model interpretability, we applied Explainable AI methods like LIME, Grad-CAM, CAM++, Score CAM and Smooth Grad to visualize the rash image regions most influential in classification. These techniques enhance both diagnostic transparency and model reliability, helping clinicians better understand the diagnostic decisions. The proposed framework demonstrates a significant advancement in automated Lyme disease detection, providing a robust and explainable AI-based diagnostic tool that can aid clinicians in improving patient outcomes.

1. Introduction

Lyme disease, an infectious condition caused by the bacterium Borrelia burgdorferi and transmitted through bites from specific ticks, has emerged as a significant global health concern. Recent studies reviewed by BMJ Global Health reveal their alarming prevalence, affecting approximately 14.50% of the global population, with particularly high rates in Central Europe (20.07%), Western Europe (13.05%), and East Asia (15.90%) [1]. The critical window for successful treatment is within 72 h post-infection [2], underscoring the urgency of early detection and intervention. This emphasizes the crucial role of medical researchers, clinicians, and professionals in the field of infectious diseases in combating this global health threat. Despite its widespread impact, Lyme disease remains one of the least publicly recognized health threats. In the United States alone, it affects an estimated 476,000 people annually [3], highlighting its rapid proliferation. Globally, between 2016 and 2019, 18 out of 252 clinically suspected cases in North India were confirmed as Lyme disease using the Standard Two-Tiered Testing Algorithm (STTA) [4]. Transmission primarily occurs through the bite of infected black-legged ticks, commonly known as deer ticks in North America, which inhabit woody and grassy areas. The characteristic bull’s-eye rash, a hallmark of Lyme disease, typically appears three to 30 days following a tick bite [5].

While various laboratory testing methods exist for diagnosing Lyme disease, the complexity and potential severity of clinical manifestations underscore the critical need for early detection. Timely and appropriate treatment can typically resolve the condition within weeks. However, challenges in early diagnosis persist, often leading to delayed interventions and increased health risks. The future of Lyme disease diagnosis looks promising with recent technological advancements, particularly in artificial intelligence and imaging. Deep learning technologies have shown promising results in accurately identifying the disease, potentially revolutionizing diagnostic approaches and improving patient outcomes. These AI-driven methods offer the potential for rapid, cost-effective, and highly accurate diagnoses, addressing the limitations of traditional diagnostic techniques. However, despite progress in applying deep learning to Lyme disease diagnosis, significant challenges remain in achieving optimal accuracy and reliability. Many current research efforts are hampered by limited or inadequate datasets that fail to capture the full spectrum of Lyme disease manifestations. Additionally, while existing deep learning models show promise, they require further optimization to enhance their performance and clinical applicability.

A notable gap in the current research landscape is the predominant focus on identifying the ticks responsible for transmitting Lyme disease, rather than analyzing and interpreting the accompanying rash images. This leaves a critical area of diagnostic potential unexplored. To address these challenges and advance the field, it is imperative to develop more refined models through the adoption of modern optimization techniques and the integration of specialized clinical expertise. This underscores the urgency for medical researchers, clinicians, and professionals in the field of infectious diseases to push the boundaries of current diagnostic methods.

Our study aims to bridge these gaps by leveraging curated datasets, optimizing deep learning models, and integrating clinical insights to improve the accuracy and reliability of Lyme disease diagnoses. We propose a novel approach combining advanced image processing techniques with state-of-the-art deep learning architectures to comprehensively analyze Lyme disease rash images. This multifaceted approach aims to enhance diagnostic accuracy and provide a more nuanced understanding of the disease’s visual manifestations. The pivotal contributions of our study are delineated as follows:

• Developed the SkinVisualNet model by combining VGG19 and DenseNet201, achieving 98.83% accuracy and 98.31% F1 score in Lyme disease rash classification which outperformed other related works that used same dataset.
• Applied gamma correction, Contrast stretching and data augmentation, boosting all model’s accuracy, Precision, recall and f1 score up to 10–13%.
• Applied LIME to provide localized, interpretable explanations for individual predictions by highlighting the specific visual features that influenced the model’s decision, thereby enhancing user understanding and trust.
• Applied Grad-CAM, Grad-CAM++, Score CAM and Smooth Grad for visual explanations by generating heatmaps that highlight important regions in input images, enhancing interpretability.

2. Related Work

Background

Lyme disease, caused by the bacterium Borrelia burgdorferi, is a tick-borne disease characterized by various clinical manifestations, making its diagnosis particularly challenging. The variability in symptom presentation among patients further complicates accurate detection and classification. Recent advancements in machine learning (ML) and deep learning (DL) techniques have opened new avenues for improving the accuracy and efficiency of Lyme disease diagnosis. This literature review synthesizes a range of studies investigating various methodologies and models for Lyme disease detection, focusing on image-based detection, biomarker identification, and genome profiling.

Shandilya and Anand (2024) utilized a publicly available Skin Disease Dataset comprising 2525 images sourced from Google, applying the EfficientNetB0 model to achieve an impressive accuracy of 98.55% [6]. Despite this high performance, the study faced challenges due to insufficient preprocessing techniques, a lack of explainable AI integration, and the absence of hybrid approaches, limiting its robustness and interpretability. Similarly, Philippe et al. [7] proposed an automated erythema migrans (EM) rash detection framework that leverages handcrafted dermatological features combined with machine-learning classifiers. Using a curated dataset of clinically confirmed Lyme disease rash images, the model achieved strong diagnostic performance, with accuracy values reaching up to 93.4%, demonstrating the feasibility of computer-assisted early Lyme disease detection. However, the study’s reliance on simple image resizing and the lack of robust cross-validation hindered its generalizability.

Mohan et al. (2025) explored the Dino V2 model on a Different Skin Disease Dataset, attaining 96.48% accuracy [8]. The study was constrained by its exclusive use of the HAM10000 dataset, class imbalance, and lack of external validation, alongside high computational demands and limited interpretability, reducing its real-world applicability. Razia et al. (2024) employed the S-MobileNet model on the HAM10000 dataset (10,000 dermatoscopic images), achieving 97.757% accuracy without segmentation and feature extraction, and 98.345% with these techniques [9]. Limitations included its restriction to HAM10000, untested generalizability, and lack of explainability (e.g., Grad-CAM), alongside challenges with threshold-based segmentation under variable image conditions.

Jerrish et al. (2023) worked with a Lyme Disease Rashes Dataset from Kaggle, originally containing 359 images (151 Lyme-positive, 208 Lyme-negative), expanded to 928 images through augmentation [10]. Using DenseNet121 with progressive resizing, they achieved 87.9% accuracy. However, the small, imbalanced dataset, reliance on augmentation, and underperformance of self-supervised models (e.g., SimCLR, MoCo) compared to CNNs, coupled with a lack of external validation and explainability, limited its clinical utility. Priyan et al. (2024) proposed a Deep Neuro-Fuzzy System (DNFS) integrating UNet, InceptionV3, XGBoost, and Mayfly Optimization, applied to a Kaggle dataset with 5059 images (4118 non-Lyme, 941 Lyme), yielding 97.36% accuracy [11]. The study’s modest dataset size, basic preprocessing, and absence of comprehensive hyperparameter tuning or real-world validation posed challenges to its deployment. Hossain et al. (2022) investigated ResNet50V2 on a Lyme disease dataset, achieving 84.42% accuracy [12]. The study was hampered by insufficient preprocessing, dataset size constraints, and lack of hyperparameter tuning. Radtke et al. (2021) applied LASSO to a dataset of 527 children with Lyme disease, achieving 95.5% accuracy [13]. However, the absence of explainable AI, hybrid models, and cross-validation limited its evaluation rigor. Dipakkumar et al. (2024) utilized a convolutional neural network (CNN) for erythema migraines detection, achieving 91% accuracy [14]. The study’s dataset lacked detailed description, and its basic preprocessing, limited model comparisons, and lack of clinical validation raised overfitting concerns.

Finally, Saravanan et al. (2023) employed an artificial neural network (ANN) on a Kaggle-sourced Lyme Disease Image Dataset of 889 images, achieving 89% accuracy [15]. The study’s reliance on basic preprocessing, lack of thorough hyperparameter tuning, and absence of external validation or interpretability methods restricted its clinical applicability.

The research suggests the importance of further exploration in the field of Lyme disease. While progress has been made, model accuracy in identifying Lyme disease remains challenging.

Table 1. Summary of Literature Review.

Author & Year	Dataset	Model Used	Limitations
G. Shandilya and V. Anand (2024) [6]	Skin Disease Dataset 2525 images publicly available to google	EfficientNetB0	Insufficient pre-processing techniques. No integration of explainable AI. No hybrid approaches.
Philippe et al. (2019) [7]	Lyme Disease 116 images taken of 63 research participants from the Mid-Atlantic region	Deep learning	Lack of robust cross-validation to ensure model generalizability. Limited preprocessing.
Mohan et al. (2025) [8]	Different Skin Disease Dataset	Dino V2	Imbalance Dataset. The model lacks interpretability. No testing was performed.
Razia et al. (2024) [9]	HAM10000 dataset, which consists of 10,000 dermatoscopic images from diverse populations	S-MobileNet	Threshold-based segmentation may struggle with variable brightness/contrast. Potential bias from class imbalance not fully addressed. Lacks explainability (e.g., Grad-CAM) for clinical trust.
Jerrish et al. (2023) [10]	Lyme Disease Rashes Dataset, containing: Original: 359 images (151 Lyme-positive, 208 Lyme-negative)	DenseNet121 with Progressive Resizing	Small and imbalanced dataset. SwAV showed high architecture sensitivity, with accuracy dropping significantly on XResNet-34. No explainability methods (like Grad-CAM) were applied, limiting transparency for clinical adoption.
Priyan et al. (2024) [11]	The study relies on a Kaggle dataset with significant class imbalance (4118 non-Lyme vs. 941 Lyme images in training)	Deep Neuro-Fuzzy System (DNFS) with UNet, InceptionV3, XGBoost, and Mayfly Optimization (MO)	No advance preprocessing applied. Comprehensive hyperparameter tuning was lacking, with only partial tuning via Mayfly Optimization. Interpretability methods such as Grad-CAM were not applied.
Hossain et al. (2022) [12]	Lyme Disease	ResNet50V2	Insufficient use of pre-processing techniques. Dataset size constraints for CNN training. Lack of hyperparameter tuning.
Radtke et al. (2021) [13]	Lyme Disease (527 Children)	LASSO	No implementation of explainable AI or hybrid model. Did not use cross-validation for model evaluation.
S. V. Dipakkumar et al. (2024) [14]	Erythema Migrans (EM) rashes (the “bull’s-eye rash” characteristic of Lyme disease)	Convolutional Neural Network (CNN)	Model comparison was limited to three architectures, excluding other strong contenders. High accuracy on a small dataset without cross-validation raises potential overfitting concerns.
Saravanan et al. (2023) [15]	Lyme Disease Image Dataset (889 images) sourced from Kaggle	Artificial Neural Network (ANN)	Pre-processing was limited to basic techniques. No thorough hyperparameter tuning was conducted for either the ANN or K-means algorithms. No interpretability methods (e.g., Grad-CAM) were used to explain predictions.

summarizes relevant studies to provide an overview of previous research efforts in this area.

3. Materials and Method

Utilizing an accurate dataset and employing an appropriate deep learning model can significantly enhance performance. We partitioned the data into training, testing, and validation sets to achieve more robust results. This process involves utilizing sub-sequence datasets. The attainment of the highest accuracy and performance by the deep learning model is essential to ensure the quality of the research. Figure 1 illustrates the overall workflow of our research.

The initial stage involves collecting images depicting Lyme disease-related skin issues and rashes. These images undergo preprocessing techniques such as gamma correction and contrast stretching to enhance their quality. Additionally, image augmentation is applied to facilitate the learning process for models, enabling them to recognize Lyme disease rash characteristics irrespective of image location or color variations.

Subsequently, models are meticulously trained on the training dataset and rigorously evaluated on both the testing and validation datasets. We employ ResNet50V2, VGG19, DenseNet201, and our proposed model for training using the collected images.

3.1. Dataset Description

The Lyme Disease Erythema Migrans Rashes dataset, sourced from Kaggle [16], comprises two distinct categories: positive and negative. The dataset includes 4118 images classified as negative (class 0) and 700 images labeled as positive (class 1). With the guidance of specialists, all redundant and duplicate images within the Lyme-positive group have been meticulously removed.

3.2. Data Preprocessing

In our preprocessing phase, we employ various techniques to enhance the quality of our data before analysis.

3.2.1. Gamma Correction

Image brightness and tristimulus values are encoded and decoded using non-linear gamma correction to accurately represent scenes with visual brightness levels aligned with human perception. This technique proves effective in enhancing object brightness differences, particularly in dimly lit images. By standardizing intensity, gamma correction improves the learning process of deep learning models, aiding in better interpretation and analysis of data [17].

$$V_{out}=A·V_{in}^{\gamma }$$

(1)

where V_out = final output image, A = constant, V_in = final input image, $\gamma$ = gamma (1.5).

3.2.2. Contrast Stretching

Image quality is enhanced through contrast stretching, also known as contrast enhancement or normalization. This process improves the grayscale dynamic range of the image. Initially, the algorithm estimates the minimum and maximum intensity levels of the image. These levels are then mapped to the desired intensity range, spanning from the lowest to the highest intensity. Subsequently, the remaining pixel intensities are adjusted using linear interpolation between these two positions. Contrast stretching enhances image visibility and standardization by increasing intensity, thereby improving feature extraction through increased contrast among neighboring pixel intensities [18].

$$I_o=\left(I_i-{Min}_i\right)\ast \left(\left({\displaystyle \frac{{Max}_o-{Min}_o}{{Max}_i-{Min}_i}}\right)+{Min}_o\right)$$

(2)

I_o = Output pixel value, I_i = Input pixel value, Min_i = Minimum pixel value in the input image, Max_i = Maximum pixel value in the input image, Min_o = Minimum pixel value in the output image, Max_o = Maximum pixel value in the output image.

3.2.3. Data Augmentation

We meticulously applied multiple data augmentation strategies to mitigate overfitting and enhance dataset diversity in our study, aiming to improve the stability and precision of Lyme disease detection models [19]. We incorporated rotations at 30, 45, 90, 180, 270, and 360 degrees to introduce variations in orientation. However, the dataset presented challenges with horizontal and vertical flips, as they resulted in duplicated image-capturing inversions. Additionally, shear transformations with factors of 0.2 and 0.5 were implemented to simulate geometric changes, which introduce distortions and challenge the model’s ability to recognize patterns. Figure 2 shows the preprocess steps and their effects.

3.3. Dataset Splitting

For model implementation we partition 80% of our data for training purposes, while the remaining 20% is reserved for testing.

3.4. Data Verification

Data validation involves randomly selecting 10 images from the dataset for quality evaluation. We employ the Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) for this purpose [20]. SSIM values range from 0 to 1, with 1 indicating perfect similarity [21]. PSNR levels above 30 dB are considered good, with a range of 20–40 dB deemed acceptable. MSE and RMSE values range from 0 to infinity, with lower values indicating higher image quality. The acceptability of these measures depends on the context and scale of images [22]. Calculating these metrics for the sample images provides a quantitative basis for assessing the dataset’s quality and reliability for analysis. In

Table 2. Value of SSIM, PSNR, MSE and RMSE.

Image	SSIM	PSNR	MSE	RMSE
	0.846624	21.174823	0.007630	0.087349
	0.931588	17.612324	0.017329	0.131639
	0.950745	21.689516	0.006777	0.082324
	0.823486	20.381293	0.009159	0.095705
	0.959904	23.746098	0.004221	0.064967
	0.938167	21.302143	0.007409	0.086078
	0.828316	20.555193	0.008800	0.093808
	0.899446	19.873140	0.010296	0.101471
	0.759238	20.984519	0.007972	0.089284
	0.909006	19.714105	0.010680	0.103346

, we present the values of image verification for 10 random images from the dataset. In this study, SSIM, PSNR, MSE, and RMSE were used solely to evaluate the effectiveness of the image-enhancement preprocessing techniques, such as, gamma correction and contrast stretching, rather than to assess the classification model itself. These metrics quantify improvements in image clarity and structural fidelity, ensuring that the enhanced images provide higher-quality visual information for subsequent feature extraction and classification.

3.5. Model Description

In our research, we carefully chose four advanced neural network architectures—ResNet50 V2, DenseNet201, VGG19, and our proprietary hybrid model, Skin-VisualNet—based on their unique architectural advantages and established effectiveness in diverse computer vision tasks.

3.5.1. VGG19

The VGG19 model is characterized by its structured design and regularity, contributing to its effectiveness. This 19-layer model utilizes 2 × 2 max-pooling for down sampling and employs small 3 × 3 receptive fields for convolutional layers. This approach is highly effective for feature extraction and hierarchical representation. Leveraging pre-trained VGG19 models from ImageNet facilitates rapid model convergence and ensures consistent performance, particularly when fine-tuned for specific applications with limited data. Initially, convolutional layers extract fundamental features, followed by max-pooling to reduce dimensionality, and finally, fully connected layers integrate these features for abstract representation [23].

3.5.2. DenseNet201

The implementation of DenseNet201 is justified by its architectural design, which enhances parameter efficiency and addresses the vanishing-gradient problem. DenseNet201 employs feed-forward dense connections, ensuring that each layer is connected to every other layer. This architecture facilitates the distribution of feature maps from previous layers to each subsequent layer, promoting feature reuse, reducing parameters, and enhancing network information and gradient flow [24]. As a result, the network can achieve greater depth while remaining efficient, as each layer contributes its learned features to the global knowledge of the network [25].

3.5.3. ResNet50V2

ResNet50V2 employs shortcut connections to skip layers, a technique that adds residual block input and output elementwise using shortcut connections. In contrast to DenseNet201, which concatenates features, this approach enables gradient backpropagation, allowing the network to learn identity functions as necessary. By doing so, it effectively mitigates the vanishing-gradient problem and facilitates the design of deeper networks without incurring performance penalties [26]. ResNet50V2 further enhances its residual blocks by reorganizing batch normalization and activation functions, resulting in improved network speed and efficiency. This optimization helps alleviate deep network optimization challenges by reducing internal covariate shifts. The model’s ability to train deeper networks without escalating computational costs or compromising performance positions it as a robust candidate for extracting nuanced visual data features [27].

3.5.4. SkinVisualNet (Proposed Hybrid Model)

In this study, we propose a novel Convolutional Neural Network (CNN) architecture that integrates the DenseNet201 and VGG19 models to create a hybrid system, aiming to enhance classification performance by leveraging the complementary strengths of both architectures. By combining VGG19 and DenseNet201 into a hybrid architecture, the resulting model can capitalize on the strengths of both networks. Specifically, the hybrid can benefit from VGG19’s high sensitivity in detecting true positives and DenseNet201’s robust and consistent feature extraction capabilities. This integration can potentially reduce the individual shortcomings of each model—such as VGG19’s tendency toward false positives—while enhancing overall classification accuracy and generalization. This makes the hybrid model a strategically sound choice for improving diagnostic performance in medical image classification tasks.

To ensure compatibility with the pre-trained DenseNet201 and VGG19 models, both of which were initially trained on the ImageNet dataset [28], we start by normalizing the input images to a size of 224 × 224 × 3. This preprocessing stage is essential to preserving uniformity and improving the standard of feature extraction.

Our hybrid model utilizes two distinct CNN architectures, DenseNet201 and VGG19, each pre-trained on the ImageNet dataset to extract features from the input images.

DenseNet201 employs dense connectivity, where each layer receives input from all preceding layers. This structure promotes feature reuse and mitigates the vanishing gradient problem, resulting in a more efficient network [29]. The output of a layer l in DenseNet201 is computed as:

$${Output}_l=H_l\left[x_0,x_1\dots \dots \dots x_{l-1}\right]$$

(3)

where $H_l$ is the transformation applied (Batch Normalization, ReLU activation, and Convolution) and $\left[x_0,x_1\dots \dots \dots x_{l-1}\right]$ represents the concatenated outputs from all preceding layers.

VGG19 consists of 19 layers with small receptive fields (3 × 3 convolutions), followed by max-pooling layers. This architecture is well-known for its depth and ability to capture intricate spatial hierarchies within images [24]. The VGG19 architecture maintains a consistent structure with multiple convolutional layers followed by pooling layers.

To reduce the dimensionality of the feature maps generated by DenseNet201 and VGG19, we apply a GlobalAveragePooling2D layer to each model’s output. This layer computes the average value of each feature map channel, transforming the high-dimensional outputs into a single vector per model. The global average pooling operation is defined as:

$$y_c={\displaystyle \frac{1}{H\times W}}{\sum }_{i=1}^H{\sum }_{j=1}^W{x}_{i,j,c}$$

(4)

where $x_{i,j,c}$ is the value of the position (i, j) in channel c, H and W are the height and width of the feature map, respectively. This operation not only reduces computational complexity but also preserves essential spatial information, making it an effective method for feature reduction [29]. Following the global average pooling, the output vectors from DenseNet201 and VGG19 are concatenated to form a comprehensive feature set. This concatenation process allows the model to combine the depth and breadth of the feature representations from both architectures, enhancing the overall feature diversity and improving classification accuracy. To stabilize the learning process and mitigate internal covariate shifts, the concatenated feature vector is passed through a Batch Normalization layer. Batch normalization normalizes the input to the next layer to have a mean of zero and a variance of one, according to the following formula:

$$\widehat{x_i}={\displaystyle \frac{x_i-{\mu }_B}{\sqrt{{\sigma }_B^2+ϵ}}}$$

(5)

where $x_i$ is the input, ${\mu }_B$ and ${\sigma }_B^2$ are the mini-batch mean and variance, and $ϵ$ is a small constant for numerical stability [30]. This normalization process adjusts and scales the activations, contributing to faster convergence and improved generalization of the model. The normalized features are then processed through a series of dense layers, each followed by ReLU activation, BatchNormalization, and Dropout. These layers are designed to refine the feature representations, with L2 regularization applied to prevent overfitting by penalizing large weights. The dense layer operation is defined as:

$$y=W^{T}x+b$$

(6)

where W is the weight matrix, x is the input, and b is the bias term. ReLU activation is applied to introduce non-linearity:

$$f\left(x\right)=\max \left(0,x\right)$$

(7)

L2 regularization is used to add a penalty to the loss function:

$$Loss = Original Loss + \lambda \sum \_iW\_i^2$$

(8)

where λ is the regularization parameter [31]. Dropout further enhances generalization by randomly setting a fraction p of the input units to zero during training:

$$y_i={\displaystyle \frac{r_i{x}_i}{p}}$$

(9)

where $r_i$ is a Bernoulli random variable and p is the probability of keeping a neuron active. The final dense layer utilizes a softmax activation function to output probabilities for each class, enabling binary classification. The softmax function converts the raw network scores into normalized probabilities:

$$Softmax(Z_i)={\displaystyle \frac{e^{z_i}}{\sum _{j=1}^N{e}^{z_j}}}$$

(10)

where $z_i$ is the input to the softmax function [32]. This ensures that the output layer provides a probabilistic interpretation of the model’s predictions. To optimize computational efficiency, we employ mixed precision training, where float16 is used for most calculations while maintaining float32 precision for critical operations. This approach reduces memory usage and accelerates training without compromising the model’s accuracy [33]. The model is compiled with the Adam optimizer, known for its adaptive learning rate capabilities, and a finely tuned learning rate of $1\times {10}^{-5}$ is selected to balance convergence speed and accuracy. The Adam optimizer updates the model weights based on the following formulas:

$$m_t={\beta }_1{m}_{t-1}+\left(1-{\beta }_1\right)g_t$$

(11)

$$v_t={\beta }_2{v}_{t-1}+\left(1-{\beta }_2\right)g_t^2$$

(12)

$${\theta }_t={\theta }_{t-1}-{\displaystyle \frac{\alpha \cdot \widehat{m_t}}{\sqrt{\widehat{v_t}}+ϵ}}$$

(13)

where $m_t$ and $v_t$ are the first and second moment estimates, $g_t$ is the gradient at time step t, ${\theta }_t$ is the parameter update, and α is the learning rate [34]. For the classification task, we employ categorical cross-entropy as the loss function, which is widely used for multi-class classification problems. This loss function measures the divergence between the true labels and the predicted probabilities:

$$L(y,\widehat{y})=-{\sum }_{i=1}^N{y}_i\mathrm{l}\mathrm{o}\mathrm{g}\left({\widehat{y}}_i\right)$$

(14)

where $y_i$ is the true label and $\widehat{y_i}$ is the predicted probability [35]. This loss guides the optimization process to improve model accuracy by minimizing the difference between the actual and predicted labels. Our proposed methodology outlines the detailed structure and components of the hybrid CNN model, integrating advanced techniques and mathematical formulations to achieve superior classification performance. The combination of DenseNet201 and VGG19 architectures, along with the use of mixed precision and regularization strategies, aims to deliver a robust and efficient solution for the target classification task. Figure 3 illustrates the layered architecture of our proposed model (SkinVisualNet).

3.5.5. Parameters of Hybrid Model

The hybrid deep learning model’s layers are described in the Parameter Table, which also includes information on the type, shape, and function of each layer. It shows important architectural elements like feature extraction from DenseNet201 and VGG19, feature concatenation, batch normalization, dense layers with L2 regularization, activation layers, and dropout layers. It also explains the flow of data through the model, from the input image tensor to the output probability tensor. The structure of the model and how each part contributes to improving the model’s performance for binary classification tasks are explained in this table.

Table 3. Parameter details of hybrid model.

Layer	Type	Trainable Parameters	Input Shape	Output Shape	Description
Input	Image Tensor	0	(224, 224, 3)	(224, 224, 3)	Input image of size 224 × 224 with 3 color channels (RGB).
DenseNet201	ConvNet	18,328,256	(224, 224, 3)	(None, 7, 7, 1920)	Pre-trained DenseNet201 model without the top layer.
DenseNet201 GAP	GlobalAveragePooling2D	0	(None, 7, 7, 1920)	(None, 1920)	Global Average Pooling layer applied to DenseNet201 model’s convolutional base.
VGG19	ConvNet	20,024,384	(224, 224, 3)	(None, 7, 7, 512)	Pre-trained VGG19 model without the top layer.
VGG19 GAP	GlobalAveragePooling2D	0	(None, 7, 7, 512)	(None, 512)	Global Average Pooling layer applied to VGG19 model’s convolutional base.
Concatenate	Concatenation	0	(None, 1920), (None, 512)	(None, 2432)	Concatenation of DenseNet201 and VGG19 outputs.
Batch Normalization 1	BatchNormalization	4864	(None, 2432)	(None, 2432)	Batch Normalization applied after concatenation.
Dense 1	Dense	1,245,184	(None, 2432)	(None, 512)	Dense layer with 512 units, applying L2 regularization.
Batch Normalization 2	BatchNormalization	2048	(None, 512)	(None, 512)	Batch Normalization applied after the first dense layer.
Activation 1	Activation	0	(None, 512)	(None, 512)	ReLU activation applied to the normalized dense layer output.
Dropout 1	Dropout	0	(None, 512)	(None, 512)	Dropout layer with a rate of 0.5 applied to the activated dense layer output.
Dense 2	Dense	131,328	(None, 512)	(None, 256)	Dense layer with 256 units, applying L2 regularization.
Batch Normalization 3	BatchNormalization	1024	(None, 256)	(None, 256)	Batch Normalization applied after the second dense layer.
Activation 2	Activation	0	(None, 256)	(None, 256)	ReLU activation applied to the normalized dense layer output.
Dropout 2	Dropout	0	(None, 256)	(None, 256)	Dropout layer with a rate of 0.5 applied to the activated second dense layer output.
Output	Dense (Softmax)	514	(None, 256)	(None, 2)	Softmax output layer providing probability distribution over 2 classes.

presents the parameters of our proposed hybrid model, detailing each component’s type, shape, and function.

4. Results

In this study, we conducted a comprehensive evaluation of four advanced neural network architectures: ResNet50 V2, DenseNet201, VGG19, and Skin-VisualNet. The performance assessment of these models was based on key metrics including precision, recall, F1 score, and overall accuracy.

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+FN+TN}}$$

(15)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(16)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(17)

Here, FP as false positive value. TP as true positive value, FN as false negative value and TN as true negative value.

$$F1-Score={\displaystyle \frac{2(recall\ast precision)}{recall+precision}}$$

(18)

4.1. Result Analysis Before Preprocessing

In this study, the performance of four deep learning models—SkinVisualNet, DenseNet201, ResNet50V2, and VGG19—was evaluated on a test set before preprocessing to assess their effectiveness in Lyme disease detection. The analysis included precision, recall, F1 score, and Area Under the Curve (AUC) metrics to provide a comprehensive understanding of each model’s performance, with overall performance ranked by F1 score to balance precision and recall, particularly for the imbalanced Positive and Negative classes.

SkinVisualNet demonstrated the overall performance with an F1 score of 41.49%. It achieved an accuracy of 85.49%, with a precision of 58.14% and a recall of 32.26%. The AUC for SkinVisualNet was 0.90, indicating strong discriminatory ability between classes. Class-wise analysis, based on its confusion matrix (781 true negatives, 36 false positives, 105 false negatives, 50 true positives), showed strong performance for the Negative class, with a precision of 0.88, recall of 0.96, and F1 score of 0.92. For the Positive class, its performance was moderate, with a precision of 0.58, recall of 0.32, and F1 score of 0.41. SkinVisualNet’s balanced metrics across both classes, combined with a high F1 score, make it a reliable model for applications requiring reasonable accuracy in identifying both Positive and Negative cases of Lyme disease.

DenseNet201 followed with an F1 score of 38.10%. It achieved an accuracy of 85.29%, with a precision of 43.14% and a recall of 34.11%. The AUC for DenseNet201 was 0.89, reflecting good classification capability. Class-wise analysis, based on its confusion matrix (785 true negatives, 58 false positives, 85 false negatives, 44 true positives), showed strong performance for the Negative class, with a precision of 0.90, recall of 0.93, and F1 score of 0.92. For the Positive class, DenseNet201’s performance was moderate, with a precision of 0.43, recall of 0.34, and F1 score of 0.38. DenseNet201’s performance indicates it is effective for the majority Negative class but faces challenges in accurately classifying the minority Positive class, though its F1 score suggests a reasonable balance compared to other models.

ResNet50V2 had an F1 score of 32.29%, placing it third in overall performance. Its accuracy was 84.47%, with a precision of 38.30% and a recall of 27.91%. The AUC for ResNet50V2 was 0.86, indicating a good but relatively lower ability to distinguish between classes compared to the other models. Class-wise, ResNet50V2 performed well for the Negative class, with a precision of 0.89, recall of 0.93, and F1 score of 0.91, as calculated from its confusion matrix (785 true negatives, 58 false positives, 93 false negatives, 36 true positives). However, its performance for the Positive class was weaker, with a precision of 0.38, recall of 0.28, and F1 score of 0.32. This discrepancy highlights ResNet50V2’s strength in identifying Negative cases but its limitation in detecting Positive cases, which is reflected in its lower F1 score.

VGG19 had the lowest overall performance with an F1 score of 8.33%, despite achieving the highest accuracy of 86.93%. Its precision was 60.00%, but it exhibited a critically low recall of 4.65%. The AUC for VGG19 was the highest at 0.91, indicating excellent discriminatory power between classes. Class-wise analysis, based on its confusion matrix (839 true negatives, 4 false positives, 123 false negatives, 6 true positives), revealed strong performance for the Negative class, with a precision of 0.87, recall of 1.00, and F1 score of 0.93. However, VGG19 struggled significantly with the Positive class, achieving a precision of 0.60, recall of 0.05, and F1 score of 0.09. This stark contrast underscores VGG19’s severe limitation in detecting Positive cases, making its low F1 score a critical drawback for applications where identifying Positive cases of Lyme disease is essential, despite its high accuracy and AUC.

Overall, the models exhibited a common trend of strong performance for the Negative class but struggled with the Positive class, likely due to class imbalance in the dataset, as evidenced by the confusion matrices (e.g., VGG19 has 843 Negative cases vs. 129 Positive cases). SkinVisualNet’s high F1 score, and balanced performance make it the most effective model for this task, followed by DenseNet201. ResNet50V2 offers moderate performance, while VGG19, despite its high accuracy, is the least effective overall due to its poor recall and F1 score for the Positive class. These results, summarized in

Table 4. Model Performance Analysis Before Preprocessing.

Model	Accuracy	Precision	Recall	F1 Score	AUC
ResNet50V2	84.47%	38.30%	27.91%	32.29%	0.86
DenseNet201	85.29%	43.14%	34.11%	38.10%	0.89
SkinVisualNet	85.49%	58.14%	32.26%	41.49%	0.90
VGG19	86.93%	60.00%	4.65%	8.33%	0.91

, and illustrated through confusion matrices in Figure 4, highlight the importance of considering balanced metrics like F1 score over accuracy alone in the context of Lyme disease detection before preprocessing. Although the full dataset contains 4818 images, the confusion matrix reflects only the test set used for final model evaluation (n = 972), as the remaining images were allocated to the training and validation sets. To illustrate the model performance before applying any pre-processing steps are shown in Table 4. The significance of applying the pre-processing steps can be found by comparing with and without adding them.

4.2. Result Analysis After Preprocessing

Our analysis highlights the superior performance of the SkinVisualNet model across all evaluated metrics, with ResNet50 V2, DenseNet201, and VGG19 also demonstrate significant effectiveness in their respective capabilities. Specifically, SkinVisualNet achieves the highest precision, recall, F1 score, and accuracy among the models, with a precision of 98.45%. The forthcoming table presents a comprehensive evaluation of model performance, including confusion matrix, precision, recall, F1 score, and accuracy metrics. Table 4 shows the model performances before the pre-processing steps were applied. Whereas,

Table 5. Model Performance Analysis.

Model	Accuracy	Precision	Recall	F1 Score
VGG19	89.43%	95.77%	92.49%	92.31%
DenseNet201	93.34%	92.98%	93.16%	93.54%
RestNet50 V2	94.92%	95.27%	95.10%	95.14%
SkinVisualNet	98.45%	99.09%	98.77%	98.83%

illustrates the results of performance measurement metrics for all models.

Figure 5 shows a comparative bar chart displaying the accuracy, F1 score, recall, and precision of four models for improved visualization.

4.2.1. Performance of the VGG19

The VGG19 model demonstrates a commendable balance between precision and recall, achieving 89.43% precision and 95.77% recall. These results highlight the model’s strong ability to correctly identify positive instances while maintaining a relatively low false-positive rate. Its overall accuracy of 92.31% further validates the model’s reliability in making accurate predictions across the dataset. Additionally, the F1 score of 92.49% indicates a well-balanced performance, effectively harmonizing both precision and recall. These figures reflect the model’s ability to maintain consistent accuracy across different classes. The results underscore VGG19’s capability to deliver a balanced and effective classification performance, as visually represented. The training and validation accuracy, alongside the training loss over 20 epochs, are presented in Figure 6, showcasing a steady convergence pattern—indicative of solid model training and minimal overfitting. Furthermore, the Area Under the Curve (AUC) was 0.98 for both Class 0 and Class 1, confirming the model’s strong discriminative power in distinguishing between classes.

4.2.2. Performance of the DenseNet201

The DenseNet201 model has demonstrated a recall rate of 92.98%, closely aligning with its F1 score of 93.16%. This practical effectiveness underscores the model’s precision of 93.34% and highlights its overall accuracy of 93.54%, affirming its quality and reliability in real-world applications. Furthermore, Figure 7 presents the training accuracy and loss curve for DenseNet201 over 20 epochs, illustrating a consistent and convergent pattern that signifies robust model training and stability. Additionally, the model achieved an AUC of 0.98 for both Class 0 and Class 1, along with a micro-average AUC of 0.98, reflecting its high discriminative capability and classification accuracy.

4.2.3. Performance of the ResNet50V2

The ResNet50 V2 model demonstrated impressive results, achieving a precision of 94.92%, the second highest among all evaluated models. Its recall rate of 95.27% and F1 score of 95.10% further highlight the model’s robust performance across classification tasks. Moreover, Figure 8 illustrates the training accuracy and loss curve for ResNet50 V2 over 20 epochs, revealing a steady and convergent pattern indicative of high training stability and performance. Additionally, the model achieved an AUC of 0.98 for Class 0, 0.99 for Class 1, and a micro-average AUC of 0.99, confirming its high discriminative power and overall classification accuracy.

4.2.4. Performance of the SkinVisualNet

The Skin-VisualNet model emerged as the top performer, showcasing exceptional classification capabilities with a precision of 98.45% and a recall of 99.09%, culminating in an impressive F1 score of 98.77%. This high recall underscores the model’s superior ability to identify nearly all relevant instances within the dataset. Furthermore, Skin-VisualNet achieved the highest overall accuracy of 98.83%, solidifying its position as the most effective model among those evaluated. This highlights the model’s remarkable precision in minimizing false positives. The accuracy and loss curves, presented in Figure 9, exhibit a smooth and convergent trend over 20 epochs, confirming that the model is well-fitted and stable throughout training. Additionally, the AUC score for Skin-VisualNet was a perfect 1.00 for both Class 0 and Class 1, as well as a micro-average AUC of 1.00, further emphasizing the model’s exceptional discriminative power and its reliability in high-stakes classification tasks. Although the model achieved a perfect AUC, indicating perfect class separability, the confusion matrix may still show non-zero misclassifications because it reflects performance at a single fixed threshold rather than across all decision thresholds.

4.2.5. Fine-Tuning and Parameter Optimization of All Model

The most important step in deep learning model optimization is fine-tuning, which involves performing precision parameter modifications to improve performance on a particular task. Our study investigates how preprocessing affects the functionality of a hybrid model that combines the VGG19 and DenseNet201 architectures. The model was tested across multiple training epochs in two different scenarios: before and after preprocessing. The results of this fine-tuning procedure are presented in detail in the following sections, with a focus on the gains in accuracy and loss metrics.

Table 6. Hyper-parameter Tuning of All Models.

Model Name	Layer	Kernel Size	Activation Function	Pooling Layer	Optimizer	Regularization	Learning Rate	Epoch Time (s)
SkinVisualNet	DenseNet201, VGG19	7 × 7, 3 × 3, 1 × 1	ReLU (Dense layers), Softmax	Global Average Pooling (GAP), None (Dense layers)	Adam	L2 (0.001), Dropout (0.5)	0.00001	5704.52
	Concatenation Layer	-	-	-		-
	BatchNormalization	-	-	-		BatchNorm
	Dense Layer 1	-	ReLU	-		L2 (0.001)
	BatchNormalization	-	-	-		BatchNorm
	Dropout (0.5)	-	-	-		Dropout (0.5)
	Dense Layer 2	-	ReLU	-		L2 (0.001)
	BatchNormalization	-	-	-		BatchNorm
	Dropout (0.5)	-	-	-		Dropout (0.5)
	Output Layer	-	Softmax	-
DenseNet201	Conv1_x	7 × 7	ReLU	MaxPooling (3 × 3, stride 2)	Adam	None	0.00001	1463.42
	Dense Block 1	1 × 1, 3 × 3	ReLU	-		BatchNorm
	Dense Block 2	1 × 1, 3 × 3	ReLU	-		BatchNorm
	Dense Block 3	1 × 1, 3 × 3	ReLU	-		BatchNorm
	Dense Block 4	1 × 1, 3 × 3	ReLU	-		BatchNorm
	Transition Layers	1 × 1	ReLU	AvgPooling (2 × 2, stride 2)		BatchNorm
	Dense Layers	-	ReLU, Softmax	GlobalAveragePooling2D		Dropout (if added)
ResNet50V2	Conv1	7 × 7	ReLU	MaxPooling (3 × 3, stride 2)	Adam	None	0.00001	1430.79
	Conv2_x Block1	1 × 1, 3 × 3, 1 × 1	ReLU	-		BatchNorm
	Conv3_x Block2–4	1 × 1, 3 × 3, 1 × 1	ReLU	-		BatchNorm
	Conv4_x Block1	1 × 1, 3 × 3, 1 × 1	ReLU	-		BatchNorm
	Conv5_x Block2–6	1 × 1, 3 × 3, 1 × 1	ReLU	-		BatchNorm
	Dense Layers	-	ReLU, Softmax	GlobalAveragePooling2D		-
VGG19	Conv1_x	3 × 3	ReLU	MaxPooling (2 × 2)	Adam		0.00001	1412.92
	Conv2_x	3 × 3	ReLU	MaxPooling (2 × 2)
	Conv3_x	3 × 3	ReLU	MaxPooling (2 × 2)
	Conv4_x	3 × 3	ReLU	MaxPooling (2 × 2)
	Conv5_x	3 × 3	ReLU	MaxPooling (2 × 2)
	Dense Layers	-	ReLU, Softmax	GlobalAveragePooling2D

presents model fine-tuning details of all models. To understand how preprocessing affects the model’s capacity to learn and generalize from the data, this comparison is essential. A clear image of the model’s development and performance gains during training is given by a selection of epochs.

The fine-tuning table shows how preprocessing significantly improves the model’s performance. These findings show that preprocessing leads to better overall performance by facilitating faster convergence, increasing learning efficiency, and strengthening the model’s capacity for generalization. The analysis highlights how crucial preprocessing is to deep learning model optimization. The increased metrics following preprocessing show that preprocessing not only speeds up the learning process but also greatly improves model accuracy and robustness. The success of integrating the DenseNet201 and VGG19 architectures inside a hybrid model framework is confirmed by this study, especially when suitable preprocessing methods are used for optimization. To further enhance model performance across a variety of datasets and applications, future research could investigate other preprocessing techniques and fine-tuning approaches.

4.3. Explainable AI

The fast growth of artificial intelligence (AI) has reshaped several sectors; nevertheless, many AI systems operate as “black boxes,” posing problems to transparency and trust. Explainable Artificial Intelligence (XAI) addresses these concerns by demonstrating how to make AI decisions interpretable and understood. To ensure the transparency and interpretability of our model’s predictions, we used Explainable AI approaches, notably LIME and Grad-CAM. These methods provide insights into our model’s decision-making processes, allowing us to trust and validate the results.

4.3.1. LIME

We have utilized Local Interpretable Model-agnostic Explanations (LIME) to elucidate our model’s predictions for the positive class (disease present). The LIME visualizations depicted in Figure 10 demonstrate the areas of the images that significantly impacted the model’s predictions.

In both images, the yellow boundary outlines the super pixels or regions that most significantly impacted the model’s classification of the positive class. These regions correspond with areas displaying visual traits typically linked to the disease, like erythema, edema, or atypical skin patterns. The LIME explanation emphasizes regions of skin inflammation, indicated by red patches, which align with symptoms related to the identified disease.

If we consider first image, the yellow contours dramatically emphasize the area of redness, indicating that the model significantly depends on this characteristic for its prediction. This aligns with domain knowledge indicating that redness is a significant significance of specific skin diseases. In the second image. The circle red lesion emphasized by LIME indicates the model’s focus on clinical features essential to the recognized disease.

4.3.2. Grad-CAM, Grad-CAM++, Score CAM and Smooth Grad Visualizations

To further interpret the model’s decision-making process for the negative class (Lyme Negative), we utilized Gradient-weighted Class Activation Mapping (Grad-CAM) and its variant Grad-CAM++ to visualize the regions in the images that most influenced the model’s classification. Figure 11 shows the original image of a skin patch labeled as “Lyme Negative” and its corresponding Grad-CAM, Grad-CAM++, Score CAM and Smooth Grad visualizations.

The Grad-CAM heatmap emphasizes the regions that significantly impacted the model’s decision-making process. The red patches, centered on the black spot and adjacent regions, indicate where the model focused its attention to generating its prediction. Despite concentrating on the area, the model classifies the image as negative, presumably due to the lack of essential Lyme disease symptoms. Grad-CAM++ provides a more detailed heatmap, emphasizing intricate components within the area. The focused analysis of detailed attributes indicates that the model employs these traits to differentiate between Lyme-positive and Lyme-negative cases, hence validating its negative categorization.

4.3.3. Implications of Applying Explainable AI

Before applying explainable AI, the model’s decision-making process remained opaque, making it difficult to assess whether its predictions were grounded in clinically meaningful image patterns or influenced by irrelevant artifacts. This “black-box” behavior limits reliability, reduces interpretability, and restricts clinical adoption. After integrating a suite of explainable AI techniques—LIME, Grad-CAM, CAM++, Score-CAM, and SmoothGrad—the internal feature-attribution process became much clearer. These methods highlighted the specific regions and pixel-level cues the model relied on when differentiating between classes, allowing us to verify consistency across multiple saliency techniques. The visual explanations revealed whether the model focused on tumor-related structures, textural patterns, or morphological boundaries, thereby increasing confidence in its outputs. Overall, explainable AI transformed the system from an opaque predictor into a more transparent and trustworthy diagnostic support tool, enabling better validation, error analysis, and alignment with clinical expectations.

4.4. Cross Validation

To ensure the robustness and generalizability of our model, we conducted 5-fold cross-validation. This approach involved dividing the dataset into five distinct subsets (folds). The model was trained on four folds and validated on the remaining fold, iterating this process five times so that each fold served as the validation set once. This method provides a comprehensive assessment of the model’s performance across different data partitions.

Table 7. Cross-validation result of 5-Fold.

Accuracy (%)	F1	F2	F3	F4	F5
Training	100	99.99	99.98	99.99	99.98
Validation	98.20	98.89	98.90	98.89	98.92

shows training and validation accuracy of every fold.

From Table 7, F1 through F5 denotes the five distinct subsets. We can see results from 5-fold cross-validation revealed that validation accuracy is between 98.20% and 98.92%. Moreover, the training and validation accuracies remained closely aligned, indicating that our model consistently delivers optimal performance without exhibiting signs of overfitting, even when subjected to varying shuffle conditions.

4.5. Comparison of Performance Before and After Preprocessing

The comparative analysis of model performance before and after preprocessing reveals significant improvements across all models. Before preprocessing, VGG19 had an accuracy of 86.93% with an F1 score of 8.33%, indicating a strong but somewhat imbalanced performance, particularly in detecting positive cases. After preprocessing, VGG19’s accuracy surged to 89.43%, with a more balanced F1 score of 92.31%, demonstrating enhanced capability in both precision and recall. DenseNet201, which had an accuracy of 85.29% and an F1 score of 38.10% before preprocessing, improved to an accuracy of 93.34% and an F1 score of 93.54% after preprocessing, reflecting a more consistent performance across both classes. ResNet50V2, initially showing an accuracy of 84.47% and an F1 score of 32.29%, saw its performance rise significantly, with an accuracy of 94.92% and an F1 score of 95.14% post-preprocessing, indicating remarkable improvements in both precision and recall. The SkinVisualNet model, which already performed well with an accuracy of 85.49% and an F1 score of 41.49% before preprocessing, emerged as the top performer after preprocessing, achieving a near-perfect accuracy of 98.45% and an F1 score of 98.83%. These results underscore the critical importance of preprocessing in optimizing deep learning models, as it significantly enhances their ability to make accurate and balanced predictions, particularly in the detection of minority classes.

Table 8. Performance Metrics Before and After Preprocessing.

Model	Precision	Recall	F1 Score	Accuracy	Accuracy	Precision	Recall	F1 Score
	After Preprocessing				Before Preprocessing
VGG19	89.43%	95.77%	92.49%	92.31%	86.93%	60.00%	4.65%	8.33%
DenseNet201	93.34%	92.98%	93.16%	93.54%	85.29%	43.14%	34.11%	38.10%
ResNet50V2	94.92%	95.27%	95.10%	95.14%	84.47%	38.30%	27.91%	32.29%
SkinVisualNet	98.45%	99.09%	98.77%	98.83%	85.49%	58.14%	32.26%	41.49%

provides a comprehensive comparative overview of performance before and after preprocessing.

4.6. Comparison of SkinVisualNet with Related Lyme Disease Detection Studies

Recent research on Lyme disease detection has explored a diverse range of data modalities and modeling strategies, reflecting the growing interest in automated early diagnosis. Laison et al. [36] leveraged global Twitter posts and emoji-based sentiment cues to identify potential Lyme cases, demonstrating the utility of social-media-driven surveillance but offering limited clinical image interpretability. Jerrish et al. [10] focused on curated Lyme rash images, employing progressive resizing and self-supervised learning to improve feature extraction, though without integrating explainability methods. Priyan et al. [11] introduced a neuro-fuzzy deep learning pipeline using UNet, Inception, and XGBoost on medical images, achieving strong accuracy with Grad-CAM visualizations to enhance interpretability. In comparison, the proposed SkinVisualNet model builds on this foundation by combining VGG19 and DenseNet201 into a hybrid architecture trained on a Kaggle-derived rash dataset, achieving higher accuracy (98.83%) while incorporating a comprehensive suite of explainability techniques (LIME, Grad-CAM, CAM++, Score-CAM, and SmoothGrad). Together, these differences highlight the contribution of this study in delivering both state-of-the-art performance and richer diagnostic transparency.

Table 9. Comparison of Lyme Disease Detection Studies.

Author (Year)	Dataset Population Focus	Dataset Explanation	Explainability Method	Accuracy (%)
Laison et al. [36] (2023)	Worldwide social media users	Self-reported Lyme-related symptom Tweets collected globally; includes textual and emoji-based sentiment features	Not applicable. (Text-based sentiment and feature-importance interpretation only)	~90% (varies by model)
Jerrish et al. [10] (2023)	Clinical skin rash images	Curated Lyme rash datasets using progressive resizing and self-supervised learning	None reported (focus primarily on SSL)	94–96%
Priyan et al. [11] (2024)	Clinical medical imaging patients	Medical image dataset processed using UNet, Inception, and a neuro-fuzzy classifier	Grad-CAM (typical for UNet-based pipelines)	97.45%
SkinVisualNet	Public Kaggle Lyme rash dataset	Kaggle-derived Lyme disease rash images with augmentation and preprocessing	LIME, Grad-CAM, CAM++, Score-CAM, SmoothGrad	98.83%

shows the comparison between some of the existing Lyme disease detection studies with SkinVisualNet, based on the dataset, population focus and explainability models implemented.

5. Discussion

5.1. Result Analysis of the Models

The results of our study underscore the critical impact of preprocessing on the automated detection of Lyme disease using deep learning techniques. Our comparative analysis of VGG19, DenseNet201, ResNet50V2, and our novel hybrid model, SkinVisualNet, provides valuable insights into how preprocessing enhances the efficacy of various architectures for this critical medical imaging task. Before preprocessing, each model exhibited reasonable performance, but there were clear limitations, particularly in terms of accuracy and the ability to consistently detect minority classes. VGG19, for example, achieved an accuracy of 86.93%, but struggled with detecting positive cases effectively, leading to a lower F1 score for this class. Similarly, DenseNet201, SkinVisualNet, and ResNet50V2 showed potential, but their performance was hampered by imbalances in the data and the absence of preprocessing techniques that could mitigate these issues. After applying preprocessing, the performance of all models improved significantly, with SkinVisualNet emerging as the top performer. This hybrid model, which synergistically combines DenseNet201 and VGG19, benefits greatly from preprocessing, which enhances feature extraction, balances class distributions, and reduces noise in the data. DenseNet201’s densely connected layers and VGG19’s deep hierarchical architecture work together to create a rich, comprehensive representation of the input data. Preprocessing further amplifies this synergy by ensuring that critical features are retained, and overfitting is minimized. SkinVisualNet, equipped with advanced techniques like global average pooling, batch normalization, L2 regularization, and dropout, demonstrates exceptional accuracy and generalization on complex medical datasets post-preprocessing. Mixed precision training accelerates computations, and the carefully tuned Adam optimizer ensures optimal training, making the model not only powerful but also efficient. The results clearly show that preprocessing is essential for achieving the highest levels of performance in deep learning models, particularly in sensitive applications like Lyme disease detection, where accuracy and generalization are paramount. In summary, the incorporation of preprocessing techniques transforms the models from reasonably effective tools into highly accurate and reliable classifiers, with SkinVisualNet leading the way in delivering superior results across all key metrics. This demonstrates the importance of preprocessing in optimizing deep learning models for medical imaging tasks, where the stakes are high, and precision is crucial.

5.2. Potentiality of Comparison with Transformer Models

Transformer-based architectures such as Vision Transformers (ViT) [37] and emerging medical-imaging Transformers [38] have shown strong performance in recent studies, their training typically requires substantially larger and more diverse datasets than those available in this work. Given the limited size and single-source nature of our dataset, incorporating such models would risk overfitting and provide an unfair comparison to the CNN-based approaches used here. Nonetheless, evaluating SkinVisualNet against modern Transformer models remains an important direction for future research as larger, clinically curated datasets become accessible.

While our study employed ImageNet-pretrained CNNs due to the limited availability of Lyme disease images, we acknowledge that pretrained natural-image features may not fully capture the subtle dermatological patterns present in medical imaging. Recent research highlights the importance of domain-specific representations, and future work will therefore focus on developing a lightweight, disease-specialized vision model trained directly on Lyme rash images and related dermatological datasets. Such a Vision Language Model, for example SmolVLM [39], would prioritize clinically relevant textures, erythema distributions, and morphological cues rather than generic natural-image features. By using parameter-efficient architectures, self-supervised learning, and few-shot adaptation, this model could be trained effectively even with small datasets while improving domain alignment and interpretability. This direction holds promise for enhancing diagnostic accuracy, reducing dependency on large pretrained models, and enabling deployment in resource-constrained clinical or mobile health settings.

5.3. Constraints of Using Single Source Dataset

The proposed SkinVisualNet framework demonstrates strong performance on the Kaggle-based Lyme disease rash dataset [16], the use of a single public repository naturally limits the diversity of patient presentations, particularly with respect to skin tone variation, imaging quality, and clinical settings. The dataset also exhibits class imbalance, which, despite mitigation strategies such as augmentation and preprocessing, may still influence the model’s generalizability. Future work will prioritize external validation using independent, larger-scale datasets that capture broader demographic and phenotypic variability, enabling more rigorous assessment of real-world robustness. Integrating comparative experiments against board-certified dermatologists will offer deeper insights into the clinical utility of the model and help position SkinVisualNet as a reliable decision-support tool in frontline diagnostic workflows. This direction will be critical for translating the current research into clinically deployable AI solutions.

5.4. Effectiveness of Pre-Processing & Post-Processing Steps

While gamma correction and contrast stretching are indeed standard preprocessing techniques, our intention was not to present them as standalone innovations but to demonstrate their combined effectiveness for Lyme rash images, which are often low-contrast and highly variable. These steps significantly improved model performance (by 10–13%) and enhanced the clarity of clinically relevant features prior to feature extraction. We have revised the manuscript to clarify that the innovation lies in the overall integrated pipeline and its demonstrated impact, rather than in the preprocessing techniques themselves.

Although the heatmaps generated by XAI methods highlight the regions most influential to the model’s predictions, their clinical meaning must be interpreted in conjunction with established medical knowledge. Visual saliency alone does not confirm disease-specific pathology; therefore, collaboration with dermatology experts is essential to validate whether highlighted regions correspond to clinically recognized features of Lyme disease rashes, such as erythema patterns, central clearing, or inflammatory borders. Integrating these medical principles ensures that the explanations move beyond technical visualization and provide clinically grounded insights that enhance trust, interpretability, and the potential for real-world diagnostic support.

5.5. Computational Complexity and Deployment Considerations

While the hybrid SkinVisualNet architecture achieves superior predictive performance, its combined use of VGG19 and DenseNet201 results in noticeably higher computational demand compared to single-stream CNN models. Both backbones individually exceed tens of millions of parameters, and their cumulative floating-point operations (FLOPs) significantly increase inference cost, particularly during high-resolution feature extraction. This elevated complexity may influence latency on CPU-only clinical workstations or mobile diagnostic devices, potentially limiting real-time applicability. Although this study prioritizes accuracy and interpretability, future work will include a full efficiency audit—quantifying total parameters, FLOPs per forward pass, and average inference time across GPU, CPU, and edge-AI hardware. Such analysis will guide model compression strategies (e.g., pruning, quantization, knowledge distillation) and inform deployment pathways tailored for resource-constrained clinical environments. Integrating these optimizations will be essential for delivering a fast, scalable, and clinically viable Lyme disease screening tool.

6. Conclusions

Our research introduces a groundbreaking methodology for disease identification through dermatological imaging, representing a significant advancement in deep learning for Lyme disease detection. The novel hybrid model, SkinVisualNet, integrates VGG19 and DenseNet201 architectures, achieving a remarkable accuracy of 98.83%. Through gamma correction, contrast stretching, data augmentation, and the integration of multiple XAI techniques—including LIME, Grad-CAM, Grad-CAM++, Score-CAM, and SmoothGrad—the model not only improved accuracy by 10–13% but also provided transparent, clinically meaningful visual explanations. This exceptional performance pushes the boundaries of medical image analysis and presents a promising tool for enhancing the accuracy and efficiency of Lyme disease diagnosis in clinical settings. This comprehensive approach enables the model to generalize across different disease manifestations, potentially leading to more reliable diagnoses in real-world clinical scenarios. The high accuracy and efficiency of SkinVisualNet could serve as a valuable tool for clinicians, augmenting their diagnostic capabilities and potentially reducing the time from initial presentation to treatment initiation. The SkinVisualNet model, with its innovative architecture and impressive performance, provides a glimpse into the future of disease detection. By building upon this foundation, we aim to contribute to a future where rapid, accurate disease detection leads to better health outcomes and improved quality of life for patients worldwide. These advancements position SkinVisualNet as a promising decision-support tool for early Lyme disease screening, with potential applications in tele dermatology, triage systems, and real-time diagnostic assistance in clinical practice.

Further research needs to corroborate the model using larger, more diverse datasets to improve generalizability. Incorporating multimodal data (clinical history, serological testing, genetic markers) could improve diagnostic precision. Improving model interpretability through attention methods or layer-wise relevance propagation might provide greater insights. So that we can overcome misclassification of other conditions with similar visual presentations, such as erythema migraine-like lesions from autoimmune disorders or other infections. Furthermore, expanding this hybrid AI methodology to additional skin diseases might enhance its influence in medical imaging. SkinVisualNet, with its unique design and outstanding performance, facilitates expedited and precise disease identification, consequently enhancing global health outcomes.