Optimized Deep Learning for Skin Cancer Diagnosis: A Salp Swarm-Optimized MobileNetV2 Approach with Grad-CAM Interpretation

The uncontrolled proliferation of aberrant skin cells is the cause of skin cancer, a major global public health concern. Although it is a complex illness with many facets, exposure to ultraviolet (UV) radiation from artificial sources, such as tanning beds or sunshine, is the most well-known contributor to the formation of malignant tumors. Biological pathways of skin carcinogenesis involve not only direct DNA damage but also immunosuppression at the local level, which is critical for tumor development and progression after UV radiation. Recent global estimates suggest that skin cancer represents a substantial portion of all malignancies, with basal cell carcinoma (BCC) and squamous cell carcinoma (SCC) being the most common types [1]. The cost of these cancers represents an important expense for health systems, and therefore, early detection accompanied by adequate diagnosis ameliorates treatment outcomes while reducing cancer-related mortalities. Cancer is a process of unrestrained tissue division, in part due to one or more hereditary transformations affecting normal cellular control [2]. This is not only very fast and with high prevalence worldwide, but it is also seen as a threat to skin cancer. It is characterized by neoplastic skin growth, leading to various types of lesions that can be benign or malignant [3]. The incidence rate of melanoma, the deadliest form of skin cancer (and which is a significant cause of death from skin disease in most developed countries), continues to rise faster than any other type and creates an urgent paradigm for public health strategies designed specifically for prevention as well as early detection. BCC [4], SCC [5], Merkel cell carcinoma [6], dermatofibroma [7], vascular lesion [8], and benign keratosis [9]. This underscores the importance of recognizing not only their unique phenotype but also the risk factors for different types, in order to develop targeted screening and intervention algorithms.

Epidermis, the outer layer of skin, accounts for most types (BCC and SCC) of Skin Cancer cases. These cancer types are most commonly caused by long-term sun exposure and usually appear on skin that is often exposed to sunlight, including the head, neck, face, arms, and hands. There is a wide geographical variation in the incidence of skin cancer, being higher where there are greater levels of sun exposure and among lighter-skinned populations. Although BCC is often restricted to a single organ and seldom invades other organs or lymph nodes, SCC is more aggressive and has the potential to intravasate nearby tissues. This highlights the importance of a timely diagnosis to prevent further damage. As a result, the term "keratinocyte cancers" is sometimes used to describe both BCC and SCC, as they arise from the most common type of skin cell, the keratinocyte [1]. Several non-invasive lesions, such as AK and Intraepithelial Carcinoma (IC), can act as precursors to SCC, showing the ability to progress if not treated [10]. This evolution emphasizes the necessity of routine skin checks and knowledge of early signs in at-risk patients.

Skin cancers are a group of diseases resulting from the uncontrolled proliferation of cells, of which UV radiation is one major cause [2]. Although melanomas account for a small percentage of skin cancer diagnoses, they are responsible for the majority of deaths. Visual examination has been developed since the 1970s to detect skin cancer, such as dermoscopy evaluation and biopsies for determining malignant from benign lesions [11]. Technological advances, such as automated image analysis and AI, have increasingly enhanced diagnostic accuracy while reducing the burden on healthcare providers. An improvement in the diagnosis of skin pathology by dermoscopy may have contributed to a rise in detection rates, making it even more important that medical practitioners are skilled at recognizing and managing cutaneous presentations effectively [12]. In conclusion, implementing a comprehensive strategy focused on skin cancer prevention, early detection, and treatment will be necessary to address this substantial public health issue.

1.1 Related work

This paper [13] examines the use of deep learning methods to develop a classifier for skin cancer, highlighting the importance of early detection. To perform this task, the study employs three different models: RegNet x006, EfficientNet v2 B0, and InceptionResNet v2, combined with a pre-processing technique that improves image quality by excluding noise, such as hairs or black pixels, in dermoscopic images. A total of four tests were carried out in order to evaluate the performance of the model on both raw and pre-processed photographs, and the final accuracy was determined to be 0.929. The results showed that pre-processing, which focuses on the region of the lesion, considerably increased the capacity to classify skin cancer in a manner that was both more accurate and sensitive.

This article [14] makes an effort to make use of deep learning techniques for the identification of melanoma, which is a kind of skin cancer that is extremely fatal. It also highlights the critical significance of making an accurate diagnosis as early as possible from the beginning. Due to the fact that conventional diagnostic procedures are time-consuming and frequently need specialized knowledge in biopsy analysis, the study focuses on how artificial intelligence might assist in making this process more efficient. A collection of 7,146 cutaneous pictures was used to assess various Convolutional Neural Network (CNN) architectures, including DenseNet201, MobileNetV2, ResNet50V2, ResNet152V2, Xception, VGG16, and VGG19. To confirm the method's performance while comparing PDFs, we ran trials on graphics processing units (GPUs) in training plots. GoogleNet was the most accurate of all the tested models, achieving a training set accuracy of 74.91% and a test rate of up to 76.08%. The study demonstrates how deep-learning models may be pivotal in improving melanoma diagnosis and assisting the clinician in making more accurate clinical decisions, further emphasizing DL's evolving role in healthcare.

Deep learning, particularly Convolutional Neural Networks (CNNs), was employed to classify skin cancer as either benign or malignant in this study [15]. The ISIC2018 dataset, comprising 3,533 skin lesion images, is utilized to enhance image quality using ESRGAN in the pre-processing step. CNN was compared with other transfer learning models, such as ResNet50, InceptionV3, and InceptionResnet, which achieved 83.2%, 83.7%, 85.8% and 84% accuracy rates. The application of ESRGAN is another excellent case for enhancing quality and classification. The study recommends future validation on larger datasets and the incorporation of other models, such as DenseNet, VGG, or AlexNet, to enhance performance.

A deep learning algorithm is used in another study [16] to categorize clinical photos of 12 different skin conditions, including melanomas, squamous cell carcinoma, and basal cell carcinoma. The CNN model Microsoft ResNet-152 was optimized and refined using 19,398 photos from the three datasets. Asan, MED-NODE, and atlas. The model was tested using independent Asan, Hallym, and Edinburgh data sets. The diagnostic accuracy of the algorithm was high, with area under the curve (AUC) values of 0.96 on the Asan dataset for basal cell carcinoma and melanoma, respectively. In contrast, the AUC values in Edinburgh were 0.90 and 0.88 for basal cell carcinoma and melanoma, respectively [13]. Using the Hallym dataset, basal cell carcinoma diagnosis sensitivity was 87.1%. Compared to 16 dermatologists, the model was validated using 480 images from Asan and Edinburgh datasets. Additional images would increase image diversity, covering a broader spectrum of ages and ethnicities.

The development of a deep learning model for 12 types of skin diseases, including malignant melanoma and basal cell carcinoma, is described in this paper [17]. This model uses the ResNet-152 architecture to distinguish among these different kinds of illness. The 3797 images were augmented 29 times by random position, scale transformation, and lightness adjustment. This enabled the project's robustness to be substantially improved. The network was then tested on 956 images, and a high accuracy of diagnosis was obtained: area under curve (AUC) 0.96 for melanoma, area under curve (AUC) 0.91 for basal cell carcinoma. With 14.1 million people around the world this year and more than 8.2 million people dying annually from skin cancer, deep learning has the potential not only to improve the accuracy of clinical judgment in detecting a lesion but also to make sure we do so as early as possible.

This paper [18] aims to enhance the clinical classification of skin conditions by introducing a unique deep-learning framework. This data set includes 13,603 dermoscopic images annotated by dermatologists and covers 14 disorders, including melanocytic nevus, psoriasis, and basal cell carcinoma. EfficientNet-b4 is the model's backbone, making it useful for the healthcare industry. The model was fine-tuned using the real-world clinical dataset, and auxiliary classifiers were incorporated to improve its performance. Visual saliency maps and t-SNE were used to study the model-learned image features. The framework achieves an accuracy (0.948), sensitivity (0.934), and specificity (0.950), all higher than that of widely used CNN models in an area under the curve (AUC) of 0.985. In a multiclass task against a panel of 280 board-certified dermatologists, this model performed comparably to them. This means that AI could potentially diagnose a wide range of skin diseases with great accuracy, engaging in both AI as research and real-world clinical care.

Using the Dermatology HAM10000 dataset, this survey [19] further explores the application of CNNs to classify seven types of malignant skin tumors, including melanoma, basal cell carcinoma, and dermatofibroma. Although CNN networks are highly successful in medical image analysis, selecting appropriate hyperparameters is challenging. Training models in several iterations is often necessary before performance can be optimized. To address this problem, the authors employed random search optimization to further optimize the hyperparameters of their CNN model, including the number of filters, kernel size, and learning rate. An initial base CNN model obtained 73.34% accuracy when developed without optimization. After random search optimization was applied, the accuracy increased to 77.17%, demonstrating the importance of such optimization. The model was deployed using Keras and TensorFlow within the Google Colab environment, pointing the way to improved skin cancer detection by refining hyperparameters in CNN networks.

This paper [20] introduces an automated framework for melanoma detection that employs deep learning technology, with a particular emphasis on utilizing pre-trained CNNs to extract features from dermoscopic images. The study arises from the branching microscopic method and has been extended to cover dermoscopy. Further challenges in melanoma classification include limited training data, similarities between classes, and variability within skin lesions. The utility of deep features derived from eight leading contemporary CNN architectures is tested against four benchmark datasets (PH, ISIC 2016, ISIC 2017, HAM10000). Results show that the DenseNet-121 combined with a multilayer perceptron (MLP) performed best, achieving an accuracy rate of 98.33% on the PH dataset and over 80% on three other sizable datasets for melanocytic lesions. It thus outperformed other models as well as existing work.

This study [21] examines the application of CNNs based on deep learning for the early and precise diagnosis of psoriasis and eczema, two common skin conditions. Acknowledging the substantial influence of prompt treatment on patient recuperation, the study employs 10-fold cross-validation to assess five cutting-edge CNN architectures. When optimized with the Adam optimizer, the Inception ResNet v2 architecture achieved a remarkable validation accuracy of 97.1%, demonstrating the model's effectiveness in diagnosing skin conditions. The research also describes two valuable applications for the developed model: a web server-oriented technique for real-time skin disease categorization and a smartphone-oriented strategy that incorporates CNN models into a mobile application. These results demonstrate how deep learning technology can improve dermatology diagnostic accessibility and accuracy.

In an effort to improve the effectiveness and precision of skin cancer diagnosis, this work [22] presents a fully automated method for multiclass skin lesion segmentation and classification. To select the most discriminant characteristics, the methodology employs an improved moth flame optimization (IMFO) algorithm, a novel Deep Saliency Segmentation method utilizing a bespoke ten-layer CNN, and local color-controlled histogram intensity values for image augmentation. The segmenting accuracies of 95.38%, 95.79%, 92.69%, and 98.70%, respectively, and the classification accuracy of 90.67% for the HAM10000 dataset are obtained when the approach is tested on the ISBI 2016 and ISBI 2017 datasets. The findings suggest that computing time is a limitation for future work and that the results are more accurate than those of existing methods, with improvements in lesion segmentation and feature optimization. Table 1 presents a summary of previous studies conducted by a group of researchers.

3.1 VGG19

The confusion matrix [39, 40] (Figure 5) of the VGG19 model gives a detailed overview of how it distinguishes whether skin lesions are malignant or benign. The model identified 351 out of the malignant cases, so it had a high sensitivity for malignancy. However, nine positive lesions were misclassified as benign, which was not ideal, considering that their false-negative rate should have been reduced in critical situations. Conversely, the model classified 184 benign cases correctly, but mistook 116 of a total of 116 Benign lesions (indicating moderate specificity). In summary, the confusion matrix highlights both the strengths and weaknesses of our VGG19 model, which still requires improvement in terms of accuracy to reduce misclassifications, particularly for malignant lesions.

Table 2. Classification report of VGG19

The classification report for the VGG19 model reveals essential performance metrics in distinguishing between malignant and benign skin lesions. As shown in Table 2, the model demonstrates a precision of 0.75 for malignant lesions, indicating a moderate capacity to avoid false positives, while exhibiting a high precision of 0.95 for benign lesions, suggesting effective identification. Notably, the recall for malignant lesions is high at 0.97, showcasing strong detection capabilities, whereas the recall for benign lesions is lower at 0.61, indicating some misclassification. The F1-scores are 0.85 for malignant and 0.75 for benign lesions, reflecting better overall performance in identifying malignant cases. With an overall accuracy of 0.81 and macro average scores of 0.85 for precision, 0.79 for recall, and 0.80 for F1-score, the report highlights the model's strengths while identifying areas for improvement, especially in benign lesion classification.

Figure 6 presents the Model Loss and Model Accuracy during training [41, 42], indicating notable trends across the epochs. The left figure shows a significant decline in both training loss (blue line) and validation loss (orange line), with the training loss sharply decreasing to around 0.4 while validation loss stabilizes slightly higher, suggesting some overfitting as the model performs better on training data than on unseen data. Conversely, the right figure depicts training accuracy gradually increasing to approximately 90%, while validation accuracy peaks at about 85%, highlighting the model's effective learning from the training set but revealing challenges in generalizing to new data. These findings emphasize the model's learning capabilities and point to areas for improving generalization.

3.2 DensNet

The confusion matrix, as shown in Figure 7, depicts the performance of a DenseNet model in classifying malignant and benign cases. Of the malignant cases, 348 were correctly predicted as malignant (true positives), while 12 were incorrectly predicted as benign (false negatives). For benign cases, 278 were correctly classified (true negatives), but 22 were misclassified as malignant (false positives). Overall, the matrix indicates a strong performance by the model, with high accuracy in both categories, although there is a slightly higher misclassification rate for benign cases.

5.png

Figure 5. Confusion matrix of VGG19

6.png

Figure 6. The training and validation loss and accuracy metrics for the VGG19

7.png

Figure 7. Confusion matrix of DenseNet

The classification report for DenseNet, Table 3, demonstrates its effectiveness in differentiating between malignant and benign skin lesions. The model achieves a precision of 0.75 for malignant lesions, indicating that 75% of its malignant predictions are correct, alongside a high recall of 0.97, meaning it accurately identifies 97% of actual malignant cases. This suggests a strong capability to minimize false negatives, which is crucial for medical diagnostics. In contrast, the precision for benign lesions is 0.95, yet the recall drops to 0.61, indicating challenges in identifying actual benign cases, which may lead to increased false negatives. The overall accuracy is 0.81, with macro averages of precision, recall, and F1-score at 0.85, 0.79, and 0.80, respectively. These results affirm DenseNet's potential as a reliable tool for skin cancer detection while indicating areas for improvement, particularly in the recall of benign classifications.

Table 3. Classification report of Densenet

3.3 ResNet results

Figure 8 Epidermal Region classification by ResNet confusion matrix. For the 360 cases of malignant records, ResNet correctly predicted 343 samples while only marking them benign for an intermediate class in more than one misclassified example. This suggests an excellent ability to identify cancer-related lesions, with many true positives. In benign cases, the model was adept at correctly classifying 279 out of a total of 300 as such, with only an incorrect classification of malignancies. The relatively small number of false positives for benign lesions also indicated that the model worked well. The confusion matrix reiterates the strong ability of ResNet to discriminate melanoma from non-melanoma skin lesions, leading to the possibility of using this classification system as a dependable diagnostic machine, especially for clinical diagnostics on practical grounds.

Table 4 presents the classification report for ResNet, demonstrating its high segmentation quality in classifying malignant and benign skin lesions. Both classes exhibit high precision, with scores of 0.94 for malignant and benign lesions, indicating low false-positive rates based on the entire dataset. Similar impressive recall values are observed at 0.95 for malignant lesions and a model-high of 0.93 for benign cases, indicating the models' high accuracy in correctly classifying actual positive cases (i.e., true positives). The averaged F1 scores, which combine precision and recall, are 0.95 for malignant lesions and 0.94 specifically for benign lesions, indicating overall performance in these metrics. The 94% macro average accuracy confirms the strength of this model, while the precision-recall and F1 scores in both classes show good classification capabilities. The results demonstrate that ResNet offers a highly effective method for classifying skin lesions and holds significant promise for deployment in clinics as a computer-aided diagnosis system for cancer detection.

Table 4. Classification report of ResNet

The training and validation curves for the ResNet model, shown in Figure 9, exhibit rapid convergence in the first two epochs, with the training loss decreasing sharply and the accuracy reaching nearly 100%. However, the validation loss and accuracy exhibit fluctuations, indicating potential instability and overfitting. While the training accuracy remains consistently high, the validation accuracy stabilizes around 90%, highlighting a gap between the model's performance on training data and its generalization to unseen data. Further tuning may be needed to improve validation stability and mitigate overfitting.

3.4 MobileNet

The confusion matrix for MobileNet, Figure 10, illustrates the model's performance in classifying malignant and benign skin lesions. The matrix shows that the model correctly classified 345 malignant and 274 benign cases. However, it misclassified 15 malignant cases as benign and 26 benign cases as malignant. This indicates that while MobileNet performs well, achieving a high level of accuracy in distinguishing between the two categories, there is still a degree of misclassification, particularly in predicting benign lesions. This error imbalance could warrant further optimization to enhance the model's sensitivity and overall classification performance.

The classification report for MobileNet, Table 5, indicates robust performance in skin lesion classification, achieving a precision of 0.93 for malignant cases and a recall of 0.96, highlighting its effectiveness in accurately identifying malignant lesions. The model had a precision of 0.95 and recall of 0.91 for benign lesions, showing that the accuracy was maintained while introducing additional false negatives. In general, MobileNet made a global accuracy performance of 0.94, with an average macro for the precision, recall, and F1-score at 0.94, indicating well-balanced skin cancer detection that may be practically feasible in real clinical applications.

8.png

Figure 8. Confusion matrix of ResNet

9.png

Figure 9. The training and validation loss and accuracy metrics for the ResNet

10.png

Figure 10. Confusion matrix of MobileNet

Table 5. Classification report of MobileNet

3.5 CNN

The confusion matrix reveals CNN classification performance concerning benign and malignant cases, as shown in Figure 11. This means that the model correctly detected 341 malignant samples (true positives) but mistakenly classified 19 benign cases as malignant (false negatives). Conversely, it accurately classified 201 benign cases, while 99 benign instances were incorrectly identified as malignant (false positives). Overall, the matrix reflects CNN's strong performance in detecting malignant cases, although some misclassifications highlight areas for potential improvement in accuracy.

11.png

Figure 11. Confusion matrix of CNN

The classification report for the CNN, Table 6, indicates a robust performance in differentiating between malignant and benign skin lesions. The model achieved a precision of 0.93 and a recall of 0.96 for malignant cases, demonstrating effective identification of malignant instances while minimizing false positives. For benign lesions, the precision was 0.95 with a recall of 0.91, reflecting reliable performance with a slight trade-off in sensitivity. An overall accuracy of 0.94, alongside consistent macro averages of 0.94 for precision, recall, and F1-score, underscores the potential utility of CNN in clinical dermatological evaluations.

Table 6. Classification report of CNN

3.6 Prediction results

Figure 12 displays two images of skin lesions and their predicted and true classes. The left image shows a lesion incorrectly predicted to be benign but actually malignant. The right image shows a lesion that was correctly predicted to be benign. These results highlight the importance of accurate skin cancer diagnosis and the potential for machine learning algorithms to assist in this process.

12.png

Figure 12. Prediction results

3.7 Results after SSO

Figure 13 shows the confusion matrix for the MobileNet model after applying SSO, demonstrating its ability to distinguish between malignant and benign cases. The model correctly identified 354 malignant instances (true positives) while misclassifying six malignant cases as benign (false negatives). In addition, it properly identified 209 benign cases, while 11 benign cases were mislabeled as cancer (false positives). Overall, this matrix shows that the MobileNet model detects malignant instances with a high accuracy rate; however, certain misclassifications indicate regions where the model's precision might be improved.

13.png

Figure 13. Confusion matrix of MobileNet after applying SSO

Table 7. Classification report of MobileNet after applying SSO

The classification report for the MobileNet model, Table 7, optimized through Salp Swarm Optimization (SSO), demonstrates excellent performance in differentiating between malignant and benign skin lesions. An accuracy of 0.93 and a recall of 0.96 for malignant cases are achieved by the model, which is able to successfully identify almost all of the true positives. It scored 0.95 precision and 0.91 recall for benign lesions. The model's overall accuracy of 0.94, along with macro averages of 0.94 for precision, recall, and F1-score, demonstrates its stability and efficacy, indicating its potential value in clinical skin cancer diagnosis.

To rigorously assess the comparative performance of the proposed MobileNetV2 model, statistical significance testing was conducted across all evaluated models. Specifically, pairwise comparisons were performed using the paired t-test based on classification accuracies obtained from multiple experimental runs. These tests were applied to determine whether the observed performance differences between MobileNetV2 (with and without Salp Swarm Optimization) and competing CNN models (e.g., VGG19, DenseNet201, ResNet50, and a custom CNN) were statistically significant. For each pairwise comparison, p-values were reported to validate whether the accuracy and F1-score improvements achieved by MobileNetV2 were not due to random variation.

Additionally, 95% confidence intervals were computed for key evaluation metrics—including accuracy, precision, recall, and F1-score—for both malignant and benign classes. For example, the SSO-optimized MobileNetV2 achieved an F1-score of 0.94 ± 0.013 for malignant lesions and 0.93 ± 0.014 for benign lesions, with an overall accuracy of 0.94 ± 0.011, indicating high consistency across folds. These confidence intervals were incorporated into the corresponding performance tables to provide clearer insights into model stability. Furthermore, a dedicated summary table was added to present the p-values resulting from the significance tests, enabling transparent, statistically grounded comparisons of MobileNetV2’s superiority over other architectures.

Table 8. Classification report of MobileNet (with 95% confidence intervals)

The two-tailed paired t-tests that were run to determine the significance of the differences in the performance of the MobileNetV2 model, which was optimized using SSO and other baseline CNN models, are displayed in Table 8. P-values with both accuracy and F1-score metrics are reported in the table in terms of pairs. The findings reveal that MobileNetV2 (SSO) had statistically significant changes ( p < 0.01), which entailed both accuracy and F1-score than VGG19 and DenseNet201 models in all the verified circumstances, which indicates its high classification ability compared to these models. Comparisons with ResNet50 (or a custom CNN) and MobileNetV2 without optimization, on the other hand, resulted in no statistically significant difference (p > 0.05), indicating that although the numerical result favors MobileNetV2 (SSO), the difference cannot be considered stable in the present experimental setup. These results support the fact that SSO is effective in increasing the performance of models and emphasize the necessity of statistical validation of the researchers when conducting comparative deep learning research. As shown in Table 9, p-values comparing MobileNetV2 (SSO) against other models (two-tailed t-test) further confirm these findings.

Table 9. p-values comparing MobileNetV2 (SSO) against other models (two-tailed t-test)

3.8 Grad-CAM

In Figure 14, the Grad-CAM (Gradient-weighted Class Activation Mapping) is visualized in a skin lesion with the predicted label of 'benign' using the MobileNetV2 model. The first row represents the initial dermoscopic picture, whereas the second one represents the Grad-CAM map of the specified picture. The heatmap reveals the sections that had the most significant impact on the model decision; the warmer colors (red/yellow) depict higher levels of relevance. It is this visual interpretability that facilitates explaining how a deep learning model makes internal decisions and enhances AI clinical diagnostic transparency.

14.png

Figure 14. Grad-CAM result

To enhance the clinical relevance of this visualization the agreement between the Grad-CAM-activated areas and those delineated by dermatologists was measured quantitatively with the Dice Similarity Coefficient (DSC). The Dice score, computed based on the presented example, was 0.87, indicating a high level of correspondence between the area of interest in the model and the ground truth lesions. Such findings indicate that the model focuses on anatomically and diagnostically significant areas, which affirm its potential reliability in real-life dermatology. Integrating these forms of quantitative validation enhances the clinical credibility of AI-driven decisions and provides a subsequent avenue for model explainability in medical imaging.

In conclusion, advances in deep learning techniques aid in the diagnosis of skin illnesses, particularly skin cancer. The suggested model takes a systematic approach to preprocessing dermoscopic images, including noise reduction, resizing, and color normalizing procedures. This may enhance image quality. A variety of CNNs, including VGG-19, DenseNet-201, ResNet-50, and MobileNet-V2, were used to classify skin lesions as malignant or benign. Salp Swarm Optimization (SSO) was used to fine-tune hyperparameters such as learning rates and batch sizes, resulting in dramatically improved diagnostic accuracy for these models. The best-performing model, MobileNetV2 with SSO, achieved 94% accuracy, a precision of 0.93 for malignant patients, and a recall of 0.96. This indicates that the model is on track for practical usage in skin disease diagnostic applications.

Several directions for future research can be explored.

First, extending the dataset to include more various skin types, patient types, and demographic features would help the model generalize and be more effective in a wider range of clinical contexts. Second, more complex optimization techniques, such as merging neural network topologies and evolutionary algorithms, could improve the model's performance.

Finally, creating real-time diagnostic tools, such as cellphones or web-based applications, would make it possible to employ these models in clinical settings. This helps dermatologists and improves patient outcomes by allowing for earlier detection and diagnosis.