Semi-supervised GAN with hybrid regularization and evolutionary hyperparameter tuning for accurate melanoma detection

This section starts with an in-depth dataset analysis, examining its characteristics and their relevance to our study. After the dataset examination, the section describes the evaluation metrics in detail. These metrics define the primary benchmarks and criteria used to assess the performance of the model. The section concludes with the presentation of experimental results. It highlights key findings and discusses their implications for the study objectives.

Datasets

This study utilized two prominent datasets to evaluate the performance of the proposed model. The first dataset originates from the SIIM-ISIC Melanoma Classification Challenge. It is included in the publicly accessible ISIC-2020 dataset, which is hosted on Kaggle. This dataset is part of the International Skin Imaging Collaboration (ISIC) Archive⁶⁷. The archive is known as the largest open-source collection of high-resolution dermoscopic images of skin lesions. The dataset includes samples from regions such as North America, Europe, and Asia. This international coverage supports the applicability of the model across diverse population groups. It comprises images collected from nearly 2,056 individuals, with 33,126 samples designated for training and 10,982 for testing. It covers a wide range of ethnic and gender groups. This diversity ensures balanced evaluation and reduces bias toward any specific population segment.

The second dataset examined in this research is HAM10000 ⁶⁸. It is recognized as one of the largest resources currently available for dermoscopic image analysis. This dataset was collected using images captured by multiple imaging devices under varied environmental conditions. This dataset was chosen because it includes many clinically relevant skin lesion types. It provides a wide variety of lesion categories that reflect real-world clinical diversity. It has also been widely adopted as a benchmark in melanoma detection studies. Its widespread use enables fair comparison with state-of-the-art models and contributes to the reproducibility and clinical validation of results. HAM10000 was compiled over 20 years through collaboration between the Department of Dermatology at the Medical University of Vienna and the Skin Cancer Clinic in Queensland, Australia. HAM10000 contains 10,015 dermoscopic images representing seven diagnostic categories. These include 6,705 samples of melanocytic nevi (nv), 1,113 of melanoma (mel), 1,099 of benign keratosis (bkl), 514 of basal cell carcinoma (bcc), 327 of actinic keratoses (akiec), 142 of vascular anomalies (vasc), and 115 of dermatofibroma (df). The dataset includes samples from both male and female individuals. It consists of 5,406 male samples and 4,552 female samples, covering different anatomical sites and age groups.

To evaluate the generalization capabilities of the proposed model, we conducted experiments using the PH2 ⁶⁹ and DermNet⁷⁰ datasets. The PH2 dataset was developed by the Dermatology Service of Pedro Hispano Hospital in Matosinhos, Portugal. It is a valuable resource for researchers and clinicians aiming to improve diagnostic precision in dermatology. This dataset al.so serves as a standardized platform for developing and validating diagnostic tools. All PH2 images were captured using the Tuebinger Mole Analyzer system. This system includes a 20× magnification lens that ensures consistency across images. The images are stored as BMP files in RGB color format, each with a resolution of 768 × 560 pixels. The dataset contains 200 images, categorized into 40 melanoma cases and 160 non-melanoma cases. DermNet is a large public dataset that contains over 23,000 expert-verified images. These images cover 23 distinct dermatological conditions, such as melanoma, eczema, psoriasis, acne, and vascular tumors.

Although four datasets are widely adopted benchmarks, it is important to acknowledge their limitations. ISIC-2020 and HAM10000 do not include explicit metadata on skin phototypes based on the Fitzpatrick scale. This limitation restricts comprehensive performance assessment across skin tones, particularly underrepresented darker types (IV-VI). The PH2 dataset follows strict imaging standards. In contrast, ISIC-2020, HAM10000, and DermNet include images taken under heterogeneous and less controlled conditions. These conditions result in variations in lighting, resolution, and imaging equipment. DermNet differs from ISIC and HAM10000 by contributing more diversity. It includes a wider range of dermatological conditions across varied skin tones and real-world photographic environments. This diversity provides useful, though indirect, insights into the behavior of the model on broader populations. Despite this improvement, the absence of standardized phototype annotation across all datasets remains a significant limitation. Future research should focus on external validation using datasets with clearly labeled skin types. These datasets should also use standardized acquisition settings to support stronger claims of generalizability.

Table 3 presents an overview of the four datasets applied in this research. To ensure a uniform classification framework across all datasets, we focus on a binary classification task (melanoma vs. non-melanoma). For datasets with multiple lesion categories, such as HAM10000, all non-melanoma lesion types are merged into a single “non-melanoma” class. This grouping ensures consistent evaluation across datasets with different labeling schemes. It also reflects a practical clinical scenario where the primary goal is to differentiate melanoma from other benign or less critical skin conditions. This approach simplifies the modeling process. It also enables the proposed SS-GAN framework to learn generalized features that are effective for melanoma versus non-melanoma tasks across diverse datasets.

The ISIC-2020 and HAM10000 datasets were fully used in this study for training, validation, and testing purposes. 70% of the samples from each class were used for training. The remaining 30% was divided equally, with 15% for validation and 15% for testing. The 70-15-15 split was selected to balance effective model training with sufficient data for tuning and testing. Using 70% of the data for training exposes the model to a wide variety of examples. This improves learning and enhances generalization capabilities. The 15% validation set helps fine-tune model parameters and reduces the risk of overfitting. The final 15% used for testing provides an unbiased method for evaluating model performance in real-world settings.

In the semi-supervised stage, we used a hybrid strategy. About 30% of the training data was labeled, while the remaining 70% was unlabeled. This approach is widely used in semi-supervised learning, as it balances the use of scarce labeled data with the abundance of unlabeled data. The labeled data provides supervision for the learning process. The unlabeled data helps the model capture additional patterns and improve generalization. This strategy improves performance on unseen data and reduces reliance on costly labeled datasets. Most studies in the field use a similar ratio, making it a standard practice in semi-supervised learning.

Before training, all input images were processed through a standardized pipeline. This ensured consistency across datasets and helped improve model performance. First, all images were resized to 224 × 224 pixels to match the input dimensions required by the convolutional backbone. Next, pixel values were normalized to the [0, 1] range. A color normalization step was also applied to reduce lighting and contrast variability across different acquisition devices. This step improves the ability of the model to generalize across diverse imaging conditions.

The HAM10000 dataset frequently contains occlusions such as hair artifacts and ruler marks. To address this issue, we applied the DullRazor algorithm. This technique detects and removes linear structures that resemble hair using morphological closing and bilinear interpolation. At the same time, it preserves important lesion details. This process reduces noise and improves lesion visibility, enhancing feature extraction during model training.

To address the class imbalance, especially in the ISIC-2020 dataset, where melanoma cases are rare, we used a hybrid strategy. This strategy combines random oversampling of the minority class with on-the-fly data augmentation. The augmentations include horizontal and vertical flips, random zooms, and slight rotations. These transformations increase diversity within underrepresented classes and reduce model bias during training. This approach ensures that the model is sufficiently exposed to all classes during training. As a result, it improves the sensitivity and robustness of the model in detecting malignant lesions.

Metrics

This article uses Accuracy, F-measure, G-means, and AUC to evaluate the proposed model. These metrics were chosen as they provide a comprehensive assessment of performance across various aspects. Accuracy measures the proportion of correct predictions but can be misleading on imbalanced datasets. F-measure, the harmonic mean of precision and recall, is useful when one class is underrepresented. G-means, the geometric mean of sensitivity and specificity, evaluates balanced performance across classes, making it suitable for imbalanced datasets. AUC, based on the ROC curve, measures the discriminative power of the model across thresholds, offering a robust and class-independent performance indicator.

In addition to these technical metrics, we use clinically significant measures, including true positive rate (TPR) and false negative rate (FNR). These metrics are especially crucial in high-risk domains such as melanoma diagnosis. TPR, also known as sensitivity or recall, measures the proportion of actual positive cases that the model correctly identifies. This is a clinically critical metric in melanoma detection, where missing a malignant case can be fatal. In such cases, a high TPR helps ensure that the model correctly identifies patients with melanoma. This makes it highly valuable in clinical decision-making. On the other hand, FNR quantifies the proportion of actual positive cases that are incorrectly classified as negative. A high FNR means the model misses many melanoma cases. This failure may result in delayed diagnosis and treatment. Such outcomes have serious consequences in real-world medical settings. As such, minimizing the FNR is crucial for ensuring patient safety.

Mathematically, Accuracy, F-measure, G-means, TPR, and FNR are defined as follows:

where

In this context, TP are cases where the model accurately identifies positive outcomes, and true negatives (TN) are cases where the model correctly predicts negative outcomes. False positives (FP) arise when the model incorrectly classifies negative instances as positive, and FN occur when the model overlooks positive instances.

Model performance

The study was conducted on a 64-bit Windows operating system using high-performance hardware that included an Intel Core i9 processor and an NVIDIA GeForce RTX 3080 GPU, which provided the computational power necessary to process large datasets and complex neural network architectures efficiently. The research utilized Python programming language with TensorFlow and PyTorch frameworks, which facilitated the development of the SS-GAN and the implementation of the improved ABC algorithm for hyperparameter tuning. The software setup was chosen to ensure robustness and speed in model training and evaluation. TensorFlow and PyTorch were selected for their comprehensive libraries, ease of use, and strong community support, which are crucial for implementing cutting-edge machine-learning techniques. GPU acceleration was critical for handling the computationally intensive tasks of training deep neural networks and processing high-resolution dermoscopic images from the ISIC-2020, HAM10000, and PH2 datasets.

We used 5-fold stratified cross-validation to evaluate the robustness and accuracy of our model on the ISIC-2020 and HAM10000 datasets. This method ensures that each fold reflects the same class distribution as the entire dataset. This is particularly important for imbalanced cases like melanoma detection, where positive cases are much fewer than negative ones. Stratification reduces model bias toward the majority class. It also improves generalization by ensuring that both melanoma and non-melanoma cases are well represented in each fold. In each round, four folds are used for training and one for testing, with all samples eventually serving as test data. This approach provides a comprehensive performance evaluation across the dataset and reduces result variance. Notably, the PH2 and DermNet datasets were used solely for external testing, not for training or validation. All reported results are presented as mean ± standard deviation (\(\:M\:\pm\:\:\sigma\:\)) to reflect consistency and reliability across folds.

During the evaluation phase, the proposed model was compared against a range of baseline methods. These included five machine learning models: MPMRF (multi-phase melanoma recognition framework)²⁴ BASCA-Opt (bat-algorithm-based skin cancer analysis optimizer)²⁵ CFBD-CM (counting fractal box dimension classification method)²⁶ NIHSC-System (near-infrared hyperspectral signal classification system)²⁷ KNN-HFIP (k-nearest neighbors hybrid-fused indoor positioning approach)²⁸.

In addition, sixteen deep learning approaches were considered: WDNR (Wavelet-based classification through deep neural architectures)⁶ECSD-Net (ensemble classification for skin disease detection network)²⁹ DDCNN-F (double decker CNN ‘F’ feature fusion)³³ YOLOv7-XAI³⁴ MRFO-SCC (manta ray foraging optimizer for skin cancer classification)³⁷ SSGNet (semi-supervised multi-path grid network for diagnosing melanoma)³⁸FixMatch-LS (semi-supervised skin lesion classification with label smoothing)³⁹ NCPLSL (noisy-consistent pseudo labeling model for semi-supervised)⁴⁰FaxMatch (multi-curriculum pseudo‐labeling for semi‐supervised medical image classification)¹³GANA-SFE (GAN-based augmentation and self-supervised feature extractor for melanoma)⁴¹ STFL (self-feedback threshold focal learning)⁴² DL-AMC (deep learning-based automated melanoma classification)⁴⁴ ODLA-Net (optimized deep learning architecture for skin lesion analysis)⁴⁵ VGG-16-GAN⁴⁶ GANViT-MD (melanoma detection via GAN synthesis and vision transformer)⁴⁷ PPO-GAN-MC (proximal policy optimized GAN for melanoma classification)⁴⁸.

Furthermore, six transfer learning models were included in the comparison: HE-TLMC (hybrid ensemble transfer learning melanoma classifier)⁵¹ DLCA-SC (deep learning comprehensive analysis for skin cancer)⁵⁵ QDLM-NM (quantized deep learning model for nail melanoma)⁵⁶ WE-TLMC (weighted ensemble-based transfer learning for melanoma classification)⁵⁹ XAI-MRA (explainable AI model for melanoma risk assessment)⁶¹ TL-MD (transfer learning-enhanced melanoma detection)⁶².

Finally, ablation studies were conducted by comparing the proposed model with five of its derivatives: Proposed w/o RL, w/o SA, w/o CR, w/o PL, and w/o HO. These versions exclude, respectively, reconstruction loss (RL), self-attention (SA), consistency regularization (CR), pseudo-labeling (PL), and hyperparameter optimization (HO) in the SS-GAN framework.

The results of these experiments for ISIC-2020 and HAM10000 datasets are summarized in Tables 4 and 5. An analysis of state-of-the-art models on ISIC-2020 and HAM10000 datasets shows that traditional ML methods underperform. They are less effective than DL and TL approaches. ML models, such as MPMRF, BASCA-Opt, and CFBD-CM, exhibit limited feature representation and poor handling of class imbalance. Their TPR values remain below 75%, with FNRs ranging from 25% to nearly 30%, indicating a weak sensitivity to melanoma. DL models offer notable improvements. MRFO-SCC, NCPLSL, and FaxMatch outperform ECSD-Net and FixMatch-LS in TPR by 10–15% and also significantly reduce FNRs. However, models like DL-AMC and ODLA-Net still struggle due to a lack of self-supervision or inadequate augmentation. TL models, such as DLCA-SC and HE-TLMC, consistently achieve TPR above 88%, benefiting from pre-trained extractors and ensemble techniques. While DL outperforms ML, TL provides additional gains through ensemble learning and explainability. However, most models still struggle to generalize to unlabeled data and remain sensitive to hyperparameter tuning in real-world diagnosis.

Across both datasets, the proposed SS-GAN outperforms all models from traditional ML, DL, and TL categories. On ISIC-2020, SS-GAN achieves a TPR of 92.825%, which is 6–12% higher than top semi-supervised models, such as STFL, FaxMatch, and GANViT-MD. Compared to the best DL model (PPO-GAN-MC), SS-GAN improves the TPR by approximately 4% and the F-measure by over 4.5%. Relative to the strongest TL model (DLCA-SC), it achieves a 4.2% increase in TPR and a 4.3% increase in F-measure. Against BASCA-Opt, it outperforms by over 19% in both metrics. On HAM10000, SS-GAN maintains an edge, with a TPR 3–10% higher than that of semi-supervised competitors and 4.9% higher than PPO-GAN-MC. It also surpasses DLCA-SC and BASCA-Opt by 2.9–13% across metrics. These results stem from the use of reconstruction loss to reduce mode collapse, self-attention for handling long-range dependencies, consistency regularization, confidence-based pseudo-labeling, and hyperparameter optimization, collectively improving feature generalization, stability, and classification fidelity.

The ablation results provide a clear justification for the architectural choices made in the proposed SS-GAN model. Each component contributes uniquely to the final performance. Removing RL results in a significant drop of over 9% in TPR and G-means for both ISIC-2020 and HAM10000. This shows its role in preventing mode collapse and maintaining latent feature consistency. Excluding SA reduces the modeling of long-range dependencies. This is reflected by a 4–6% decrease in TPR in ISIC-2020 and a slightly smaller drop in HAM10000. Omitting CR decreases stability in the presence of input perturbations. This results in a 3–5% performance drop, particularly evident in the ISIC-2020 dataset. The PL mechanism enhances the use of unlabeled data by filtering low-confidence predictions; its absence leads to TPR and AUC drops of up to 8% in HAM10000. HO ensures convergence to optimal configurations; without it, the model underperforms by 6–7% across key metrics in both datasets. Notably, prior semi-supervised models such as FixMatch-LS and GANViT-MD did not integrate all these components, particularly lacking CR and HO strategies. Therefore, the combined effect of all architectural elements in SS-GAN explains its superior generalization, robustness, and accuracy across diverse imaging scenarios.

Two-tailed paired t-tests were performed on six metrics (Accuracy, F-measure, G-means, AUC, TPR, and FNR) to validate the statistical significance of the improvements, comparing SS-GAN with the top models. On ISIC-2020, SS-GAN significantly outperformed BASCA-Opt with p-values of 0.0004, 0.0006, 0.0002, 0.0009, 0.0011, and 0.0013, respectively. Against PPO-GAN-MC, values ranged from 0.0015 to 0.0026, while for DLCA-SC, they were between 0.0019 and 0.0042. For semi-supervised models like FaxMatch and STFL, p-values were all below 0.003. On HAM10000, p-values compared to BASCA-Opt were 0.0003 to 0.0012, versus PPO-GAN-MC 0.0021 to 0.0040, and versus DLCA-SC 0.0028 to 0.0046. Overall, comparisons between the proposed model and other state-of-the-art models across both datasets show consistently low p-values. 95% confidence intervals remained narrow (± 0.01 to ± 0.05) across both datasets. These findings show that the proposed model surpasses existing models in performance metrics. It achieves this with high statistical significance, ensuring that the observed improvements are both authentic and meaningful.

Table 6 Provides an analysis of the computational efficiency of various models. It assesses runtime and GPU usage across the ISIC-2020 and HAM10000 datasets. The proposed SS-GAN exhibits favorable runtime (3042s on ISIC-2020, 2698 s on HAM10000) and moderate GPU usage (21.2 GB, 18.2 GB). Compared to semi-supervised baselines like FixMatch-LS and faxmatch, SS-GAN improves runtime by 8–18%. It also reduces GPU usage by up to 35%, averaged across both datasets. Furthermore, SS-GAN is more efficient than the top DL model, PPO-GAN-MC (3512s, 28.5GB on ISIC-2020). It also surpasses the best TL model, DLCA-SC (3267s, 28.4GB on ISIC-2020). On ISIC-2020, SS-GAN reduces runtime by 13.4% compared to PPO-GAN-MC and by 2.8% compared to DLCA-SC. It also lowers GPU usage by 25.6% and 25.4%, respectively. Similar trends are observed on HAM10000. These findings validate SS-GAN as a scalable, resource-efficient solution well-suited for clinical deployment.

Figure 4shows the training and validation loss curves for 250 epochs on the ISIC-2020 and HAM10000 datasets. In both cases, the proposed model demonstrates stable convergence, with training and validation losses decreasing consistently. The minimal gap between the curves indicates strong generalization and the absence of overfitting. Especially on the HAM10000 dataset, the nearly overlapping loss curves highlight robust learning with minimal variance. On ISIC-2020, some fluctuations are present. However, the overall downward trend confirms effective convergence. These results validate the reliability and learning efficiency of the model across datasets.

Figure 5 illustrates the accuracy trends of the proposed SS-GAN in comparison with baseline SSL models, including FaxMatch (feature-based augmentation for semi-supervised learning), MixMatch (a comprehensive semi-supervised learning approach), ReMixMatch (improved MixMatch for semi-supervised learning), Π-model (pi-model), and UDA (unsupervised data augmentation). The comparison is performed on the ISIC-2020 and HAM10000 datasets across varying ratios of labeled data. SS-GAN consistently achieves superior accuracy, maintaining a 4–8% margin even with only 20% of the data labeled. This advantage stems from its architecture. Reconstruction loss mitigates mode collapse and increases data diversity, improving coverage of rare melanoma subtypes. Self-attention captures global spatial dependencies in dermoscopic images, ensuring critical features are not overlooked. Pseudo-labeling enhances the training set by incorporating high-confidence unlabeled samples, while consistency regularization stabilizes predictions in the presence of perturbations. Hyperparameter tuning with ML-ABC ensures optimal learning dynamics. It also prevents overfitting and achieves smooth convergence in challenging SSL settings. SS-GAN maintains its advantage as labeled data increases, proving scalability and adaptability. It outperforms other SSL methods by efficiently leveraging both labeled and unlabeled data. Its success comes from robust regularization and tuning. These mechanisms ensure consistent improvements and reliable performance, even under low-label conditions or varying labeling ratios.

Figure 6 illustrates the decision-making time distributions of the proposed model in real-time settings, evaluated on the ISIC-2020 and HAM10000 datasets. The results show that most predictions were made within 150–190 ms for ISIC-2020. For HAM10000, prediction times ranged from 135 to 185 ms, confirming real-time suitability across both datasets. This demonstrates that the model consistently maintains low latency, a crucial requirement for real-time clinical settings. The tight distribution and low average inference time confirm the suitability of the mode for time-sensitive applications. It is well-suited for automated skin lesion screening, where rapid decision-making is critical for workflow efficiency.

Figure 7 illustrates instances of misclassification by the proposed SS-GAN model in the HAM10000 dataset. These errors highlight the need for further refinement of classification models, especially for cases with subtle differences between melanoma and non-melanoma lesions. Despite the strong performance of the model, these misclassifications show where it struggles. Distinguishing rare melanoma cases from non-melanoma lesions becomes harder under complex imaging conditions. This analysis highlights the need for enhanced training to improve generalization across diverse datasets and effectively handle the nuances of skin lesion classification.

Figure 8shows the ROC and precision-recall (PR) curves of the proposed model for the ISIC-2020 and HAM10000 datasets. The AUC values of 0.898 (ISIC-2020) and 0.908 (HAM10000), along with PR-AUC scores of 0.775 and 0.821, respectively, indicate excellent discriminatory power. High PR-AUC values indicate that the model maintains strong precision across a wide range of recalls. This is critical for imbalanced classification problems. These curves confirm that the proposed model can reliably distinguish minority classes from dominant ones.

Figure 9 presents the confusion matrices for the ISIC-2020 and HAM10000 datasets, demonstrating the strong performance of the proposed model. The model achieves a high number of TN and TP. It records only 29 FN and 557 FP for ISIC-2020, and 11 FN and 104 FP for HAM10000. These low FN and FP rates indicate that the model is reliable in identifying melanoma cases while minimizing misclassifications. This makes it highly suitable for diagnostic applications.

Figure 10 shows SHAP-based visualizations that highlight the key regions influencing the decisions of the model. In each image, areas shaded in red and yellow indicate features that significantly contribute to the final classification, while purple and blue signify regions that make a low contribution. Board-certified dermatologists have verified these high-importance areas as clinically relevant and linked to malignancy or atypical features. These visualizations demonstrate that the model consistently focuses on lesion boundaries and high-risk pigment zones, suggesting effective feature attribution. Consistent patterns across lesion types confirm the interpretability and transparency of the model. This supports its clinical reliability in dermatological diagnostics.

Analysis of generalizability

We evaluated the generalizability of the proposed model by conducting experiments on the PH2 and DermNet datasets. These datasets provided additional context to assess their performance compared to existing models. Importantly, all samples from both PH2 and DermNet were used exclusively for evaluation purposes. No additional training or fine-tuning was performed on these datasets. Instead, we directly applied the model trained solely on the ISIC-2020 dataset. This zero-shot evaluation protocol provides a rigorous and unbiased method for testing the generalization of the model across domains and skin-type distributions, thereby avoiding dataset-specific adaptation.

The results are presented in Tables 7 and 8 for the PH2 and DermNet datasets. As seen in Table 7, the proposed model significantly outperforms all compared ML, DL, TL, and SSL methods on the PH2 dataset. The proposed method improves the F-measure by + 24.9%, G-means by + 29.8%, and TPR by + 21.9% compared to BASCA-Opt. Even compared to strong DL models like SSGNet and FaxMatch, F-measure improves by + 6.8% and + 11.5%. Compared to the best TL model (DLCA-SC), it achieves a 5.9% increase in F-measure, a 4.3% increase in G-means, and a 4.9% increase in TPR. These results highlight the robustness of the learned features beyond the distribution of the training data.

Similarly, Table 7 confirms the consistent superiority of our model on the DermNet dataset. The proposed model surpasses the best classical ML model (BASCA-Opt) by + 21.3% in F-measure, + 21.8% in G-means, and + 21.3% in TPR. Against strong SSL models, such as FixMatch-LS, NCPLSL, and GANA-SFE, the proposed SS-GAN achieves F-measure improvements of + 14.7% to + 18.6% and TPR improvements of + 14.2% to + 15.9%. Compared to the best TL model (DLCA-SC), the proposed model achieves a 5.4% increase in F-measure, a 4.4% increase in G-means, and a 4.4% increase in TPR.

These results show that the proposed SS-GAN performs exceptionally well across multiple external datasets without retraining. This confirms its strong generalization ability, robust feature extraction, and domain adaptability, which are key requirements for real-world dermatological diagnostic applications.

We performed two-tailed paired t-tests on six metrics (Accuracy, F-measure, G-means, AUC, TPR, FNR) to compare the proposed SS-GAN with top models on PH2 and DermNet datasets. On PH2, SS-GAN outperformed BASCA-Opt with p-values ranging from 0.0004 to 0.0009. Against PPO-GAN-MC, p-values ranged from 0.0012 to 0.0021. For DLCA-SC, values ranged from 0.0017 to 0.0028. On DermNet, SS-GAN showed stronger results compared to BASCA-Opt, with p-values ranging from 0.0002 to 0.0006. Against PPO-GAN-MC, p-values ranged from 0.0011 to 0.0017. For DLCA-SC, values ranged from 0.0019 to 0.0027. Overall, comparisons with state-of-the-art models on PH2 and DermNet show similarly low p-values. Confidence intervals (95%) were ± 0.01–0.08. These findings demonstrate that the proposed model surpasses existing models with high statistical significance. They confirm that the observed improvements are authentic and meaningful.

Analysis of SS-GAN

This section compares the proposed SS-GAN with various GAN variants selected for their relevance and effectiveness in medical imaging or generative tasks. These include style-based GAN version 3 (StyleGAN3)⁷¹ deep regret analytic GAN (DRAGAN)⁷² adversarial generator-encoder (AGE)⁷³ alpha GAN (α-GAN)⁷⁴ GAN, diffusion GAN (DGAN)⁷⁵ melanoma high-fidelity GAN (MELIIGAN)⁵⁰ super-resolution GAN (SRGAN)⁷⁶ enhanced super-resolution GAN (ESRGAN)⁷⁷ efficient-GAN (EGAN)⁷⁸ StarSRGAN⁷⁹ and the original SS-GAN. StyleGAN3 is well-known for generating high-quality, diverse image features, making it a benchmark in medical image synthesis. DRAGAN and AGE improve training stability and help prevent mode collapse, which is critical for medical datasets with class imbalance. α-GAN and DGAN are considered advanced hybrid GAN architectures that combine multiple generative and discriminative strategies. SRGAN, ESRGAN, and StarSRGAN have shown strong performance in medical image super-resolution. MELIIGAN is a recent GAN tailored for melanoma image fidelity, making it highly relevant for our domain.

We evaluate the performance of the proposed SS-GAN using four metrics: Maximum Mean Discrepancy (MMD), Kullback–Leibler Divergence (KLD), Wasserstein Distance (WD), and Mode Score (MS). These metrics assess diversity (mode coverage) and fidelity (realism) of generated samples. Both aspects are crucial for understanding mode collapse. MMD measures the distance between real and generated data distributions in a reproducing kernel Hilbert space. A lower MMD indicates better alignment and improved diversity. KLD quantifies divergence from the real distribution and is sensitive to missing modes. WD measures geometric distance between distributions, providing meaningful gradients and insights into fidelity. MS combines the number of distinct modes with discriminator confidence. It evaluates how well the generator produces varied yet realistic outputs.

To ensure fairness, all other model components were standardized across tests. The performance outcomes are detailed in Tables 9 and 10 for the ISIC-2020 and HAM10000 datasets. The proposed SS-GAN shows substantial improvements in both fidelity and diversity compared to leading GAN models, particularly StyleGAN3, MELIIGAN, and the baseline SS-GAN. On the ISIC-2020 dataset, it achieves an MMD of 0.264. This is 87% lower than MELIIGAN (2.053) and 94% lower than StyleGAN3 (4.511), indicating superior alignment with real data distributions. The KLD of 0.855 is 76% lower than MELIIGAN (3.626) and 84% lower than StyleGAN3 (5.202). Similarly, the WD of 0.877 is 79% lower than MELIIGAN (4.317) and 87% lower than StyleGAN3 (6.929), confirming improved realism. For diversity, the MS reaches 0.827, which is 21% higher than MELIIGAN (0.682) and over 55% better than SS-GAN (0.531). Similar trends are observed on HAM10000, with a 65% reduction in MMD and a 79% lower KLD compared to MELIIGAN, as well as nearly a 90% improvement over StyleGAN3. These results confirm the superior performance of SS-GAN.

We performed two-tailed paired t-tests on four GAN-specific metrics (MMD, KLD, WD, and MS) to evaluate the statistical significance of the proposed SS-GAN compared to other GAN models. On ISIC-2020, SS-GAN significantly outperforms MELIIGAN with p-values of 0.0006 (MMD), 0.0008 (KLD), 0.0004 (WD), and 0.0010 (MS). Comparisons with StyleGAN3 yield even lower p-values: 0.0003, 0.0005, 0.0002, and 0.0009, confirming the strong improvements. Confidence intervals are narrow (± 0.01 to ± 0.03), validating the findings at a 95% confidence level. On HAM10000, SS-GAN again surpasses MELIIGAN with p-values of 0.0004, 0.0007, 0.0005, and 0.0009, and StyleGAN3 with 0.0002, 0.0004, 0.0003, and 0.0007. Confidence intervals remain tight (± 0.01 to ± 0.04) at the same confidence level. Overall, comparisons with state-of-the-art GAN architectures on both datasets yield low p-values and narrow confidence bounds. These results reinforce the reliability of our enhancements and confirm that the proposed SS-GAN generates diverse, realistic, and generalizable melanoma samples.

Figure 10 compares the data distributions of the original and proposed SS-GAN with real data for the ISIC-2020 and HAM10000 datasets. The proposed SS-GAN closely aligns with the real data distribution, effectively avoiding the mode collapse seen in the original SS-GAN, which concentrates around fewer data modes. This improvement comes from integrating reconstruction loss, self-attention mechanisms, and consistency regularization. These components enhance feature diversity and stabilize the training process. As a result, the proposed model captures a wider range of data variations and preserves the multimodal structure. This ensures the generation of realistic and representative samples compared to conventional GAN models.

Figure 11 visualizes feature vectors generated by the model generator for the ISIC-2020 and HAM10000 datasets. These outputs are originally high-dimensional. Direct visualization of these high-dimensional vectors is difficult, so PCA was applied to reduce them to two dimensions for clearer interpretation. The left plot illustrates the distribution of generated features with reconstruction loss, showing wide dispersion that reflects improved diversity and reduced mode collapse. In contrast, the right plot (without reconstruction loss) reveals dense and localized clusters, indicating limited variability. This comparison confirms that reconstruction loss mitigates mode collapse, allowing the generator to create richer feature representations, which are essential for robust melanoma classification.

Analysis of the proposed ML-ABC

This section compares the performance of ML-ABC with several popular hyperparameter tuning methods. The analysis considers three basic methods: random search (RS), grid search (GS), Bayesian optimization (BO), and Hyperband (HB). It also evaluates eleven evolutionary algorithms, including human mental search (HMS), salp swarm algorithm (SSA), cuckoo optimization algorithm (COA), firefly algorithm (FA), bat algorithm (BA), particle swarm optimization (PSO), deferential evolution (DE), CMA-ES, puma optimizer (PO)⁸⁰ arithmetic optimization algorithm (AOA)⁸¹ and the original ABC.

To ensure fair assessment, uniformity was maintained across all model variables during evaluations. The results from the ISIC-2020 and HAM10000 datasets are detailed in Tables 10 and 11. Compared to BO, ML-ABC improves accuracy by 7.6% on ISIC-2020 (90.52% vs. 84.08%) and 6.6% on HAM10000 (92.36% vs. 86.63%). Against the top evolutionary algorithm, AOA, ML-ABC shows gains of 10% on ISIC-2020 (90.52% vs. 80.38%) and 12% on HAM10000 (92.36% vs. 82.18%). Compared to the original ABC, ML-ABC boosts accuracy by 18.9% on ISIC-2020 and 18.1% on HAM10000. F-measure and G-means show similar improvements. These gains result from the ability of ML-ABC to share knowledge among sub-populations, accelerate convergence, and avoid local optima. Overall, ML-ABC demonstrates superior stability and generalization compared to all baselines.

We conducted two-tailed paired t-tests on six metrics (Accuracy, F-measure, G-means, AUC, TPR, FNR) to establish statistical significance. The proposed ML-ABC was compared with BO, the strongest basic optimizer, and the AOA, the leading evolutionary optimizer. On the ISIC-2020 dataset, ML-ABC outperformed BO with p-values ranging from 0.0015 to 0.0023 across all metrics. Comparisons with AOA showed p-values of less than 0.003. A similar trend was observed on the HAM10000 dataset, where p-values ranged from 0.0015 to 0.0024 against BO and remained below 0.0035 against AOA, confirming robustness. Confidence intervals for ML-ABC were narrow (± 0.02–±0.04 at a 95% confidence level), confirming its stability and reliability. Overall, each comparison shows that ML-ABC achieves faster convergence, robust exploration, and superior performance compared to other optimization techniques.

Figure 13 shows the loss minimization process over 300 iterations for ISIC-2020 and HAM10000 using the proposed ML-ABC method. Both curves display a smooth and steady decline, confirming the strong convergence capability of the algorithm. The HAM10000 dataset exhibits a slightly faster reduction in loss, while ISIC-2020 shows consistent stability across iterations. This result demonstrates the effectiveness of the mutual learning component in ML-ABC, which improves convergence and enhances hyperparameter tuning.

Figure 14 illustrates the effectiveness of the ML-ABC algorithm in optimizing key generator hyperparameters for the ISIC-2020 and HAM10000 datasets. The figure indicates that the algorithm identifies optimal values for batch size, number of epochs, learning rate, and MLP layers, achieving maximum accuracy for each dataset. Notably, the curves display clear peaks at the optimal configurations, confirming the sensitivity of the model to tuning. These results demonstrate that ML-ABC effectively explores the search space, avoids local optima, and identifies hyperparameter settings that maximize performance.