Skin diseases pose significant challenges in healthcare because of their diverse appearances and the need for accurate, specific diagnoses. Traditional diagnostic methods rely heavily on dermatologists' expertise, which can be subjective, time-consuming, and resource-intensive. These challenges have driven the increasing adoption of deep learning techniques to automate skin disease classification, aiming to enhance diagnostic accuracy and efficiency. To address these challenges, numerous studies have explored the use of deep learning techniques for automating skin disease classification.

Several studies have explored the application of deep learning in skin disease classification. For instance, De et al. (2024) automatically identified skin diseases using dermatoscopic imaging. The authors also mentioned that, traditionally, dermatologists manually examined pigmented skin lesions, which can be subjective and time-consuming. Agarwal and Godavarthi (2023) evaluated and compared various skin diseases in terms of cosmetics and common skin issues. The author's dataset included ~25,000 records for the eight most common skin conditions, and a convolutional neural network was used to achieve imaging performance comparable to or superior to that of humans. Naeem et al. (2024) proposed SNC_Net, which applies deep learning models that include features from dermoscopic images and handmade (HC) feature extraction approaches to improve classifier performance. Alshahrani et al. (2024) combined CNN models (DenseNet121, MobileNet, and VGG19) with Random Forest (Rf) and Feed Forward Neural Networks (FFNN) networks to obtain complex features from dermoscopy images, and the result was a hybrid system capable of early detection of various skin lesions. Behara et al. (2024) employed adaptive thresholding to extract regions of interest (ROI) and improved cancer detection accuracy through dynamic capabilities. Khan et al. (2021) employed the PAD-UFES-20 dataset, which includes six different types of imbalanced skin cancer types, to address the data imbalance using data augmentation. Wei et al. (2023) presented a convolutional neural network model for skin disease categorization using model fusion. Sharma et al. (2023) analyzed the HAM10000 dataset to assess the profitability of several Convolutional Neural Network (CNN) designs in concealing seven different types of skin lesions. In this study, we employed EfficientNets, which outperforms standard designs because of its lightweight design. Venugopal et al. (2023) implemented improved EfficientNet B4 and EfficientNet V2-M models to categorize malignant and benign skin lesions in dermoscopic images. Adegun and Viriri (2020) proposed a framework to segment and classify skin lesions to automatically detect skin cancer. The suggested structure of this problem is divided into two parts. The basic part of this network is used to analyze mixed and complex data problems. In this step, the algorithm acquires the surrounding details using decryption. Ravi (2022) proposed classifying and detecting learning-based mixed features grouped for skin cancer that consider attention costs. Shimu et al. (2022) proposed the implementation of transfer learning to six forms of skin diseases: peeling, acne, eczema, heat rash, melanoma, and cold sores. The skin conditions were classified using a Convolutional Neural Network. Sadik et al. (2023) focused on employing CNNs to become adept at skin disease recognition, similar to the renowned detective Sherlock Holmes. They wanted to know how well these architectures could perform the job. AlSuwaidan (2023) compared six popular CNNs (VGG16, EfficientNet, InceptionV3, MobileNet, NasNet, and ResNet50) to determine which one was the best at guessing the top three skin issues in the Middle East. They cleaned up the images a bit with some filtering and denoising to ensure the models had the best view. Jaisakthi et al. (2023) developed a Deep Convolutional Neural Network (DCNN)-based model to cartegorize skin cancer into melanoma and non-melanoma. Milantev et al. (2020) conducted the ISIC Skin Lesion Classification Challenge using dermoscopic images and patient information to study skin lesions. Tahir et al. (2023) developed DSCC_Net, which uses a CNN to classify skin cancer and evaluates it on three big datasets (ISIC 2020, HAM10000, and DermIS). Gupta et al. (2020) presented a system that employs transfer learning with pretrained models to improve the results. Furthermore, this study used a CNN to precisely detect and categorize skin cancer. Pham et al. (2020) suggested a mixed strategy to address class disparities in skin disease categorization. This process combined data-level balanced mini-batch logic with real-time image augmentation and algorithm-level generation of new loss functions. Abbasi et al. (2024) proposed a modified VGG16-based algorithm to distinguish real and AI-generated medical images. The model was trained and fine-tuned using hyperparameter tuning on a dataset of 10,000 synthetic skin lesion images produced by a GAN. It distinguished real images from AI-generated images with 99.82% accuracy. Gamage et al. (2023) trained an Xception-based model for melanoma classification using the HAM10000 dataset. Bayesian hyperparameter optimization was employed alongside Grad-CAM/Grad-CAM++ heatmaps for improved explainability, where the model reached an accuracy of 90.24% while providing critical input regions that influence model predictions.

Table 1 summarizes the key contributions, findings, and limitations of recent studies. These studies highlight the significant progress made in applying deep learning to skin disease classification while also identifying areas for improvement, such as addressing overfitting, improving data diversity, and ensuring computational efficiency.

Following a thorough review of existing skin disease datasets, we selected a public resource that provides extensive coverage of 19 skin conditions (Skin Diseases Dataset, 2024),. including Vitiligo, Psoriasis, Acne, Actinic Carcinoma, Atopic Dermatitis, Cellulitis, Eczema, Drug Eruptions, Herpes HPV, Light Diseases, Lupus, Melanoma, Poison Ivy, Benign Tumors, Systemic Disease, Ringworm, Urticarial Hives, Vascular Tumors, Vasculitis, and Viral Infections. Each condition was represented by sufficient images for both training and testing, thereby enhancing the dataset's applicability to dermatological research. Figure 2 shows sample images from the dataset, demonstrating its diversity.

We employed a systematic and targeted data augmentation strategy to enhance the robustness and generalizability of the dataset. This approach improved the diversity and representativeness of the dataset while addressing class imbalances in underrepresented categories.

Image augmentations included random 30-degree rotations, zooms in or out by 20%, and 10%-20% horizontal or vertical shifts. These augmentations represented the common real-world variation in image posture, scale, and location, facilitating a broadly varied dataset that accurately depicts realistic conditions (Ayoub et al., 2023). We also applied horizontal flipping to add more variance. These augmentation techniques were implemented by utilizing TensorFlow's ImageDataGenerator to implement real-time augmentation during the preparation of the dataset. This not only maintained the original image quality but also allowed for uniform expansion of the dataset.

Particular focus was given to classes where the number of original images was limited, specifically classes with 5, 16, 37, 26, 29, and 46 samples. For these classes, the augmentation process was repeated numerous times by using different combinations of transformations to accrete a much larger dataset. This allowed a more proportionate representation of each class, reducing the bias and creating a more robust diversity in the overall dataset. This ensured that the dataset was broadening its representativeness of clinical variability found in the real world by generating more samples for rare conditions, which made analyses performed later much more robust. Table 2 presents the number of images per class before and after augmentation and divides the augmented dataset into training, validation, and testing subsets based on a 70:15:15 ratio.

To standardize the dataset for consistent model input, several preprocessing steps were applied. Images were resized to a uniform 128 × 128 pixels, and color values were adjusted to ensure consistency across samples. Pixel values were normalized to a 0–1 scale, improving training stability and convergence speed. The labels were one-hot encoded to format them appropriately for classification tasks, thereby enabling the model to distinguish between different classes effectively (Ayoub et al., 2022).

For model evaluation, the dataset was divided into training, validation, and testing sets, with 70% of the data allocated to training and 15% each for validation and testing. This balanced split supports robust model training, validation, and testing, ensuring generalizability across unseen data.

The proposed model for skin disease prediction combines the strengths of DenseNet121 and EfficientNetB0 through a hybrid transfer learning approach. Each architecture was selected for its unique capabilities, resulting in a robust model that could handle the complexity and variability of skin disease images.

where H_i represents a composite function of operations such as batch normalization, ReLU activation, and convolution, and x₀, x₁, x_l−1 is the concatenated output of all previous layers. This dense connectivity ensures efficient feature reuse and mitigates the vanishing gradient problem, enabling the network to learn fine-grained features that are essential for distinguishing between similar skin conditions.

where φ is a user-defined scaling coefficient, and α, β, and γ are constants determined via a grid search. This allows EfficientNetB0 to maintain a manageable size while capturing high-level features effectively.

By combining DenseNet121 and EfficientNetB0, the hybrid model leverages DenseNet121′s capability for detailed pattern recognition and EfficientNetB0′s efficiency, resulting in a scalable and accurate architecture suitable for real-time applications.

The proposed model was developed using Python 3.8, OpenCV 4.7, and the Keras library 2.8 for model building and image processing. The development environment was run on Windows 10 Pro with the following hardware configuration: Intel i5 processor (2.9 GHz), Nvidia RTX 2060 GPU, and 16 GB RAM.

To evaluate the model's performance, we used standard evaluation metrics, including accuracy, precision, recall, and the F1 score. These metrics provide a comprehensive assessment of the model's predictive accuracy and robustness, which is particularly important for multi-class classification tasks in skin disease prediction.

In this section, we evaluate the performance of the proposed hybrid model for skin disease prediction using training and validation accuracy and training and validation loss. These metrics provide insights into the model's learning process, generalizability, and robustness against unseen data.

The training and validation accuracy (see Figure 4) shows a consistent increase over the 100 epochs, with the training accuracy at the last epoch reaching ~98.18%. The high training accuracy indicates that the model was able to learn complex functions that perfectly predicted the level of skin disease class from the training dataset. The validation accuracy stabilized at ~97.57%, indicating that the model generalized well on new data. The similarity between training and validation accuracy suggests that the model has not overfitted because it learned specific features of the training data while being able to perform well on images it has never seen before.

This may indicate that a specific hybrid architecture and training approach fit the given dataset well. Overall, the increasing trend of training vs. validation accuracy indicates that the model is robust and suitable for deployment. Moreover, these results emphasize the usefulness of the dropout and selective layer freezing techniques to handle overfitting.

Trends on training and validation loss curves shown in Figure 5 validate the behaviors observed for model performance. The training loss is consistently decreasing and converging to low levels, indicating successful optimization of model parameters. The validation loss also decreased, but with a few fluctuations, suggesting that the models were learning well while not overfitting significantly. The validation loss approximates the training loss in the last epoch, which indicates stable model performance and good generalizability.

These observations validate the robustness of the hybrid model architecture and the training strategies employed, notably the dropout and selective freezing of the layers. The small difference between the training loss and validation loss indicates that the model does not overfit to unseen data.

The hybrid architecture, which combines DenseNet121′s fine-grained feature extraction with EfficientNetB0′s efficient scaling, clearly contributed to the model's high performance. The fluctuations observed in the validation accuracy and loss curves, although minimal, can be attributed to the variability in the dataset. This variability is especially common in medical imaging, in which similar visual features may appear across different classes (e.g., rashes and discoloration in multiple skin conditions). The use of dropout and selective layer freezing helped mitigate these fluctuations, thereby supporting the model's ability to generalize without compromising its learning depth.

To further assess the model's performance, we evaluated its predictions on the test set, which provided a realistic measure of its ability to generalize to unseen data. The confusion matrix for the test set (see Figure 6) provides a detailed view of how well the model performed across each skin disease class, with an overall accuracy of 97.57%. The high test accuracy underscores the robustness of the model because it consistently achieved correct classifications across most classes.

The confusion matrix shown in Figure 6 demonstrates strong predictive power for the majority of classes with minimal misclassifications. Each class generally had a misclassification rate between 2.2 and 2.5%, suggesting that the model was capable of distinguishing between different skin disease categories with high precision. The low misclassification rate also indicates that the model successfully learned the nuanced features needed to separate visually similar skin conditions, which is particularly challenging in dermatology.

Within the matrix, certain classes, such as Melanoma, Psoriasis, and Vitiligo, showed very high precision and recall, as indicated by the nearly full diagonal dominance, reflecting the model's effectiveness in identifying the unique features of these conditions. However, a few classes with overlapping visual symptoms, such as Eczema and Dermatitis, experienced minor misclassifications. These errors, which generally involve one to three images per class, may stem from shared visual characteristics such as texture or color patterns. Such subtle misclassifications suggest that further data augmentation or fine-tuning can enhance the model's differentiation ability in these overlapping categories.

To evaluate the performance of the hybrid model, we compared it to state-of-the-art (SOTA) models such as DenseNet121, EfficientNetB0, VGG19, MobileNetV2, and AlexNet. As shown in Table 3, this comparison demonstrates that the proposed hybrid train-validation approach exhibits better performance for all three metrics. The adaptive hybrid model performs better than other architectures by leveraging DenseNet121′s ability to learn minimal/complex features along with EfficientB0 runtime performance, yielding an architecture with good accuracy and runtime performance for medical images.

The use of transfer learning and extra dense layers allows the model to perform well on a small dataset. As transfer learning uses pretrained weights, the model can utilize the features learned earlier. Not only does this speed up training, but the proposed method also significantly improves the network's ability to learn complex representations and thereby increases accuracy. By integrating these techniques—a hybrid architecture to handle high-dimensional input data, transfer learning to leverage pretrained models, and a robust model design with performance-driven features—we established the hybrid model as a top-performing skin disease classifier, outperforming conventional architectures.

Although the hybrid model takes 5 h 14 min 37 s to be trained, it is a fair trade-off, especially considering VGG19, which trained for 8 h 15 min 32 s but has drastically lower performance (85% accuracy vs. 98.18%). Furthermore, its training time is only slightly worse than the standalone DenseNet121 or EfficientNetB0, regardless of having implemented both architectures and obtaining a significant jump in performance. This combination of DenseNet121 and EfficientNetB0 enabled the hybrid model to attain an outstanding accuracy rate (98.18%) along with remarkable precision (0.95), recall (0.96), and F1 Score (0.95), thus marking a revolution in skin disease classification.

The hybrid model also took 22 s to process 2,889 images during the testing, which is only a slight delay compared to lightweight models like MobileNetV2 or AlexNet. However, the slight increase in inference time was compensated by the near-optimal classification performance. Faster models that trade off large amounts of computation for time with little accuracy include AlexNet and MobileNetV2, both of which report significantly reduced accuracy on data such as medical images (80 and 83%, respectively) and are unable to handle much more complex data that may be found in the medical field. As evident from the performance metrics of the hybrid model, it has a greater ability to generalize and therefore is more reliable option for applications in high-stakes environments like clinical diagnosis.

The detailed results in Table 3 demonstrate that the hybrid model achieves an optimal trade-off between computational complexity and performance, thereby rendering it the most competitive model among the compared models. This approach is recognized as the best and most practical approach for medical image diagnosis due to its ability to provide excellent results without restricting itself from achieving predictions on bigger data, thereby providing scalable results.

Furthermore, the proposed model was benchmarked against several existing models in the literature (Table 4). The comparison includes key performance metrics such as recall, precision, and accuracy, which are essential for evaluating the effectiveness of classification models. Notably, the proposed model's performance is on par with or outperformes the majority of the existing models.

From the table, it can be observed that the hybrid method proposed in the last row achieves the highest recall (0.9600) and precision (0.9500), although its accuracy (0.9818) is slightly lower than that of Naeem et al.'s model (0.9691). The models trained on ImageNet, such as Calderón et al. (2021), Jain et al. (2021), and Naeem et al. (2022), demonstrate competitive results in accuracy and recall, with Naeem et al.'s model achieving the highest accuracy. Interestingly, the model by Saarela and Geogieva (2022), which does not mention a pre-training method, has the lowest performance in terms of accuracy (0.8000), highlighting the importance of pre-training for enhancing model performance. In comparison to these models, the proposed model's performance in terms of recall, precision, and accuracy positions it as a strong contender in the field, showcasing the effectiveness of the model's architecture and the choice of pre-training techniques. The higher recall and precision values in the proposed model suggest better classification performance, particularly in terms of accurately distinguishing between classes and minimizing false positives.