Accurate classification of skin lesions, particularly for melanoma detection, remains a critical challenge in medical image analysis. Leveraging recent advances in deep learning, this paper investigates the use of curriculum learning in ensemble deep learning context for melanoma classification. To validate the proposition, three primary strategies are compared: transfer learning of CNNs using VGG16, ResNet50, and EficientNetB0 models; ensemble learning techniques such as bagging; and curriculum learning that progressively guides training in increasing order of complexity. Experiments conducted on the ISIC 2019 and 2020 dermoscopic image datasets demonstrate that curriculum learning applied to EficientNetB0 achieves superior classification performance, reaching an F1- score of 90.77%, outperforming conventional fine-tuning and ensemble approaches. These results underscore the potential of integrating curriculum learning in ensemble learning context with state-of-the-art CNN architectures to improve the robustness and accuracy of automated melanoma diagnosis.
The accurate and early diagnosis of skin cancer, particularly melanoma, is a critical public health
challenge due to its aggressive nature and potential for metastasis if undetected. Traditional manual
interpretation of dermatological images by clinicians is a complex, time-consuming task prone to
inter-observer variability, misdiagnosis, or delayed treatment [
Artificial Intelligence (AI), and specifically deep learning (DL), has emerged as a transformative force
in medical imaging, ofering unprecedented opportunities to enhance diagnostic eficiency and accuracy
[
Despite these advancements, several challenges persist in deploying DL models for reliable clinical diagnosis. Models frequently struggle with pronounced class imbalance in medical datasets, high intraand inter-class variability of lesions, and the presence of ambiguous or atypical cases that lead to uncertain predictions [6]. While ensemble learning methods [7], which aggregate predictions from multiple models, can mitigate these issues by improving predictive stability and leveraging model diversity [8], they do not address the fundamental learning process on challenging samples. To foster more robust representation learning and address diagnostic uncertainty, curriculum learning (CL) has attracted increasing attention [9]. Inspired by human cognitive development, CL involves structuring the training process by progressively introducing data samples from simple to complex, thereby guiding the model’s optimization and enhancing its generalization capabilities [10].
This work systematically investigates advanced deep learning strategies to enhance the automated classification of skin lesions, with a particular emphasis on demonstrating the eficiency of curriculum learning in this challenging domain. Our core objective is to show that a confidence-based curriculum used in ensemble deep learning context, when applied to powerful pre-trained CNNs, can surpass the performance of both conventional TL and ensemble methods for melanoma detection. The main contributions of this paper are summarized as follows: • A comprehensive empirical evaluation of five pre-trained CNN architectures (VGG16, ResNet50, InceptionV3, EficientNetB0, and Xception) under two distinct transfer learning modes: feature extraction and fine-tuning. • The implementation and assessment of a bagging-based ensemble approach applied to the EficientNetB0 architecture, designed to enhance predictive stability and robustness for skin lesion classification. • The development of a novel curriculum-learning framework in ensemble deep learning context, where training data is dynamically ordered by diagnostic dificulty (derived from initial classifier confidence), enabling a progressive and more efective model adaptation process. • A rigorous experimental study conducted on a two-class dataset derived from the ISIC 2019 and ISIC 2020 challenges, providing a robust benchmark for binary classification of melanoma versus nevus.
The remainder of this paper is structured as follows: Section 2 provides an overview of related work in automated skin lesion classification. Details of the proposed methods are then provided in section 3, including the transfer learning strategies, the ensemble approach, and the curriculum learning approach applied in ensemble deep learning context. In section 4, the experimental results are presented and discussed, comparing the performance of all implemented approaches. Finally, Section 5 concludes the paper and outlines promising directions for future research.
In line with numerous research fields and applications influenced by AI revolution over the past decade, the automatic detection and classification of skin cancer from medical images has also witnessed significant advancements. In this context, several automated systems using AI models and techniques have been introduced to facilitate early non-invasive diagnosis. This section provides a review of prior work and existing methodologies for skin cancer detection and classification, spanning from traditional machine learning techniques to state-of-the-art deep learning and ensemble approaches, highlighting their evolution, key contributions, and limitations.
Early eforts in automated skin cancer detection primarily used traditional machine learning algorithms, which typically relied on handcrafted features (such as texture, color, and shape descriptors) extracted from images, followed by classification (like SVM, KNN, and decision trees). While providing preliminary solutions, their main limitation was the inability to automatically capture the complex discriminative features inherent in dermatoscopic images, often requiring tedious expert intervention.
An early work by Codella et al. (2015) [11] proposed a method that combined deep learning with sparse coding and SVM for melanoma recognition. Aiming to reduce the reliance on extensive annotated data, the authors tried to benefit from unsupervised learning and transferring features from natural photographs, mimicking expert clinical reasoning. Evaluated on 2,624 clinical cases, their method achieved 93.1% precision for distinguishing melanoma from non-melanoma lesions. More recently, Wei et al. (2024) [12] explored the eficacy of classical machine learning algorithms for early skin cancer detection. They applied SVM, KNN, and Decision Tree algorithms on a large ISIC dataset (53,177 images across seven classes). Their results indicated that KNN achieved the highest multiclass precision at 74.93%, while SVM ofered better binary classification with an AUC of 0.676. Despite their interpretability and lower computational demands, these methods often require expert-driven feature engineering and struggle to capture complex image patterns, limiting their diagnostic accuracy and scalability. Consequently, these approaches remain valuable as baseline references but are increasingly being outperformed by deep learning models capable of autonomously learning highly discriminative features.
The significant potential of CNNs was underscored by Esteva et al (2017) [
To address the limitations of standard CNNs, researchers have introduced some enhanced and hybrid
deep learning architectures. These models often combine the strengths of diferent network types or
incorporate task-specific mechanisms. Catal Reis et al. (2022) [ 15] introduced InSiNet, a CNN adapted
for skin cancer classification, achieving up to 94.59% precision on ISIC datasets and outperforming
conventional deep learning and traditional machine learning methods. In a comparable way, Shah et
al. (2024) [16] integrated explainable AI with Xception for early melanoma detection, attaining 98.5%
precision on ISIC 2018 and enhancing interpretability via Grad-CAM and LIME. Building on the idea of
specialized components, Kavitha et al. (2024)[
To address issues like model bias, variance, and limited generalizability, ensemble methods were explored where predictions of multiple individual models aggregated to produce a more robust and accurate ifnal decision. These techniques frequently outperforming any single model. In fact, numerous studies, in the context of medical image classification, have demonstrated the eficiency of such techniques. For instance, Khaled et al. (2023) [18] proposed a stacking framework that combines multiple CNN architectures to improve classification accuracy across diverse medical imaging modalities and anatomical regions. In a related efort, for brain lesion detection, Laribi et al. (2024) [ 19] demonstrated that ensembles combining CNNs and ViTs achieved significant improvements. More recently, Khaled et al. (2025) [20] integrated progressive transfer learning with ensemble methods to achieve significant enhancements in mammogram-based breast cancer diagnosis. In the context of skin lesion analysis, ensembles techniques have likewise shown strong potential. Rahman et al. (2021) [21] demonstrated substantial gains, improving recall to 94% by aggregating five diverse CNN architectures on HAM10000 and ISIC datasets for classifying seven types of skin lesions. Bassel et al. (2022) [22] extended this with multi-level stacking of CNN-derived features and traditional classifiers (SVM, Random Forest, KNN, and logistic regression). This yielded 90.9% precision and F1-score, highlighting the benefit of cross-paradigm integration, albeit at increased computational cost. Moreover, subsequent work has focused on enhancing ensemble pipelines through data enrichment and optimization. Chang et al. (2022)[23] combined MELA-CNN image features with clinical metadata (age, sex) and applied K-means SMOTE to address class imbalance, achieving an F1-score of 86.1% and an AUC of 0.970 with XGBoost. Shortly after, Thanka et al. (2023)[24] integrated VGG16-based features with XGBoost and LightGBM classifiers, leveraging GAN-generated synthetic samples to balance the ISIC data and reach 99.1% of precision and 99.4% of recall, though reliance on synthetic data may risk overfitting. Natha et al. (2024)[25] further advanced ensemble strategies by combining Max Voting across pre-trained ensemble models (Random Forest, Gradient Boosting) with Genetic Algorithm-based feature selection, achieving 95.80% of precision and 95.20% of F1-score on ISIC 2018 and HAM10000, while reducing overfitting and improving generalization. Ghosh et al. (2024)[8] extended the feature diversity concept by combining deep learning embeddings from VGG-19, Capsule Network, and ViT with multiple machine learning classifiers through majority voting, achieving 91.6% precision. In addition, Gamil et al. (2024) [ 26] integrated EficientNet-B0 features with PCA and an AdaBoost and SVM ensemble, attaining 93% precision on DermIS and 91% on ISIC, indicating strong diagnostic potential despite persistent interpretability challenges. While ensemble methods significantly boost robustness and accuracy, they often come with increased computational complexity and can reduce the overall interpretability of the diagnostic decision.
In this work, we investigate a range of deep learning strategies for the binary classification of skin lesions, specifically diferentiating melanoma from nevus. The methodological framework is designed to enhance both predictive accuracy and model robustness, leveraging advanced CNN architectures and contemporary training paradigms. For this, we structured the study around three approaches: (i) transfer learning, (ii) ensemble deep learning via bagging, and (iii) curriculum learning used in ensemble deep learning context. A comprehensive description of the dataset is first provided, followed by a detailed explanation of each proposed approach.
Our experiments were conducted on the ISIC 2019 and 2020 Melanoma Dataset [27], a fusion of the ISIC 2019 [28, 29, 30] and ISIC 2020 [31] challenges. The ISIC 2019 comprises over 25,000 dermoscopic images across nine lesion categories, including 4,522 melanoma cases, while the ISIC 2020 dataset contains 33,126 images across two classes with only 584 melanoma cases. For our study, only two diagnostic categories of high clinical relevance were considered, namely melanoma (MEL) and nevus (NV) resulting in 11.449 images (5,106 MEL, 6,343 NV). To mitigate class imbalance, we applied random under-sampling of the majority class. Figure 1 presents representative example images from each class.
Given their documented success in medical image analysis, five CNN architectures (VGG16, InceptionV3, Xception, ResNet50, and EficientNetB0) were employed as backbone models in this approach. To leverage the advantages of transfer learning, each model was initialized with weights pre-trained on the ImageNet dataset. Two transfer learning strategies were explored, resulting in two distinct experimental scenarios for this approach: feature extraction and fine-tuning.
In the feature extraction scenario, as illustrated in Figure 2, the convolutional backbone was frozen, and only the classification layers were trained, thereby utilizing the pretrained weights exclusively for feature representation while adapting the classifier to the target task. Conversely, in the fine-tuning scenario (illustrated in figure 3), a subset of the upper convolutional layers was unfrozen and jointly optimized with the classification head, enabling the adaptation of high-level feature representations to the specific characteristics of the dermatological domain. This approach seeks to achieve an optimal trade-of between the generalization benefits of pretrained representations and the specialization required for accurate skin lesion classification.
Convolution Backbone (example: VGG16) 56 x 56 x 256 28 x 28 x 512
In this approach, Ensemble methods were explored. To enhance predictive robustness and reduce variance, we employed an ensemble learning strategy based on bootstrap aggregating (bagging) using multiple EficientNetB0 models. In this framework, multiple instances of EficientNetB0 are each trained (fine-tuned) on distinct bootstrap-resampled subsets of the original training set, ensuring diversity among base learners. During inference, predictions from all trained models are combined through majority voting to yield the final classification output. The detailed procedural steps of this approach are outlined in Algorithm 1.
Input: Training dataset = {(train, train)}, number of ensemble members , base model (EficientNetB0) Output: Final prediction ^
Initialization: // set of trained base models
Inspired by human pedagogy or learning behavior, the curriculum learning (CL) process begins by relying on simpler samples and progresses to more complex data to improve generalization, reduce sensitivity to noisy or ambiguous samples, and facilitate more stable convergence.
In this approach, we propose a curriculum-learning pipeline that leverages the EficientNetB0 model as a backbone to be fine-tuned, due to its favorable trade-of between accuracy and computational eficiency. The following steps detail the process of this approach (visualized in figure 4), starting from ranking the samples by dificulty to the final validation phase: 3.4.1. Dificulty-Based Ranking This stage aims to identify and rank skin lesion images according to their diagnostic complexity. The output of this step is an ordered dataset, starting with the most straightforward cases (characterized by distinct morphological patterns, absence of artifacts, and minimal noise) and progressing to the most challenging cases, involving subtle lesion features, imaging artifacts, or significant anatomical variability. This ordered structure is obtained by leveraging the confidence scores produced by the model’s softmax activation, which serve as a quantitative measure of predictive certainty, thereby enabling a systematic and reproducible arrangement of images from easiest to most dificult. 3.4.2. Progressive Curriculum Construction To allow for a gradual learning, we need to partition the ordered dataset into subsets, each corresponding to a level of dificulty, which will be used to train the model gradually or progressively. Each image is thus associated with a certainty level derived from the predicted probability of its ground-truth label. In our approach, we chose to partition the training set into five such dificulty levels, defined in Table 1: 3.4.3. Training phase The training process begins with the least complex images, enabling the model to acquire fundamental visual representations and recognize clear diagnostic patterns. Performance is then periodically assessed on a validation set, enabling the refinement of model parameters before progressing to succeeding higher-complexity cases. Subsequently, increasingly challenging images are introduced in a progressive manner by training on added complexity level subsets and validating accordingly. This allows the model to incrementally enhance its feature representations and improve its ability to address atypical or ambiguous cases. 3.4.4. Final consolidation across the entire dataset After completing or running over all complexity levels, the model is trained on the full training dataset to solidify learned representations and reinforce its generalization capability. 3.4.5. Final validation on an independent test set Finally, for a rigorous assessment of its robustness and diagnostic accuracy. The model is evaluated on an unseen dataset (test set) that reflects real-world clinical scenarios.
BAG 1 BAG 2 BAG 3 BAG 4 BAG 5
Our experiments involved training and evaluating five ImageNet-pretrained CNN backbones (VGG16, InceptionV3, Xception, ResNet50, and EficientNetB0) using diferent training strategies on a balanced subset of the ISIC 2019 and ISIC 2020 datasets. To enhance generalization and mitigate overfitting, we applied data augmentation techniques including random flips, brightness/contrast/saturation adjustments, and random crops, alongside stratified train/validation/test splits. Models were trained for 25 epochs with an early stopping patience of 5, and curriculum learning was employed by gradually progressing through five dificulty levels before consolidation on the full dataset. Performance was evaluated using Accuracy, Precision, Recall, and F1-score. Typically, the F1-score is the performance indicator of choice when it comes to comparative evaluations of multiple approaches in medical data analysis, as this clinical context is particularly sensitive to false positives and false negatives. For this reason, the F1-score was adopted as the primary metric in our study.
As outlined before, we evaluated three deep learning approaches for the binary classification of skin lesions, distinguishing melanoma from nevus on a combined dataset from ISIC 2019 and ISIC 2020. The results of these evaluations are presented in the following subsections.
In this approach, we trained and evaluated five CNN architecture (InceptionV3, Xception, VGG16, ResNet50, and EficientNetB0) applied under transfer learning paradigms, both as feature extractors (scenario1) and with fine-tuning (scenario 2). Initial experiments of Scenario 1 where transfer learning was applied with frozen convolutional layers revealed varied performance across the diferent CNN architectures. ResNet50 and EficientNetB0 achieved the highest F1-scores among the models, as illustrated in Table 2, with ResNet50 reaching an F1-score of 88.79% and EficientNetB0 closely following at 88.66% . VGG16 also performed competitively with an F1-score of 87.07%, while InceptionV3 and Xception lagged behind with scores of 81.60% and 84.55%, respectively. These results indicate that the architectures, such as EficientNetB0 and ResNet50, are better suited for capturing discriminative features in dermoscopic images.
To improve robustness and reduce variance, we implemented ensemble learning using bagging with EficientNetB0 as the base learner. The ensemble aggregated predictions from multiple independently trained instances of EficientNetB0 models. This approach yielded consistent improvement achieving an F1-score of 88.86% as reported in Table 4.
The most significant performance gains were observed when applying curriculum-learning strategy combined with fine-tuned EficientNetB0 model. Curriculum learning involved structuring the training process to present images in increasing order of complexity, allowing the model to progressively learn from simpler to more challenging examples. This approach led to a marked improvement in classification metrics, with EficientNetB0 achieving an F1-score of 90.77%, as listed in Table 5, surpassing both the single-model transfer learning and ensemble bagging approaches.
The results of this study highlight the significant benefits of applying curriculum learning strategy for the classification of skin lesions using deep convolutional neural networks, particularly EficientNetB0. Curriculum learning, by organizing training samples from easier to harder based on a confidence-derived dificulty metric, appears to facilitate a more efective learning process. This progressive exposure enables the model to first capture robust feature representations from clear, unambiguous examples before adapting to more challenging, noisy images. Such a training strategy likely contributes to improved convergence stability and ultimately yielded a higher F1-score (90.77%) compared to other approaches. Furthermore, the curriculum learning method not only improved the F1-score but also enhanced precision and recall, reaching 93.74% and 87.99%, respectively, indicating a better balance between false positives and false negatives, which is critical in clinical contexts . This suggests that curriculum learning efectively guides the model to better generalize by focusing on easier samples ifrst, thereby stabilizing the learning process and reducing misclassification rates. On the other hand, the ensemble learning via bagging revealed complementary strengths compared to curriculum learning, evidenced by an F1-score of 88.86% for EficientNetB0 ensembles, highlighting the validity of this approach in reducing the variance of the predictions and improving robustness of the model. As well, regarding transfer learning approaches, fine-tuning enhanced performance across most CNN models compared to pure feature extraction, with notable gains observed for VGG16. This aligns with existing literature emphasizing the importance of adapting pretrained weights to domain-specific data for improved representation learning. Nonetheless, curriculum learning surpassed fine-tuning alone, delivering the largest absolute improvement in classification performance within this experimental framework. In summary, the integration of curriculum learning into the training pipeline for skin lesion classification ofers a robust and eficient means to improve model accuracy and stability beyond traditional transfer learning and ensemble methods. These findings underscore the value of incorporating data-driven sample dificulty ordering in deep learning workflows, particularly in medical domains where data complexity and variability pose significant challenges. Future work should explore the generalizability of curriculum learning across diverse datasets and investigate its synergy with other advanced techniques and modalities to further enhance clinical applicability.
This work presented a comparative study of transfer learning, ensemble deep learning, and curriculum learning strategies for melanoma classification from dermoscopic and clinical images. Our findings confirm the efectiveness of deep CNNs in skin lesion analysis and provide evidence that curriculum learning ofers tangible advantages over conventional or even ensemble methods. By gradually exposing models to increasingly dificult samples, curriculum learning used in ensemble deep learning context improves feature refinement and enhances robustness when faced with atypical or ambiguous lesions. Beyond performance gains, curriculum learning contributes to better generalization, an essential property for clinical deployment where unseen cases may deviate significantly from training data. The integration of such progressive training schemes into automated diagnosis systems could therefore support dermatologists with more reliable and eficient decision-making. Future work will focus on expanding curriculum design by incorporating alternative dificulty measures, such as low-level image quality indicators or complexity metrics derived from auxiliary models. Additionally, evaluating the approach across larger and more heterogeneous datasets will be critical to validate its applicability in real-world clinical environments.
[5] L. Alzubaidi, M. Al-Amidie, A. Al-Asadi, A. J. Humaidi, O. Al Shamma, and M. A. Fadhel, “Novel transfer learning approach for medical imaging with limited labeled data,” Cancers, vol. 13, no. 7, p. 1590, 2021, doi: 10.3390/cancers13071590. [6] S. Nahavandi, M. Abdar, M. Samami, S. D. Mahmoodabad, T. Doan, B. Mazoure, R. Hashemifesharaki, L. Liu, A. Khosravi, U. R. Acharya, and V. Makarenkov, “Uncertainty quantification in skin cancer classification using three-way decision-based Bayesian deep learning,” Comput. Biol.
Med., vol. 135, p. 104418, 2021, doi: 10.1016/j.compbiomed.2021.104418. [7] D. Müller, I. Soto-Rey, and F. Kramer, “An analysis on ensemble learning optimized medical image classification with deep convolutional neural networks,” IEEE Access, vol. 10, pp. 66467–66480, 2022, doi: 10.1109/ACCESS.2022.3182399. [8] S. Ghosh, S. Dhar, R. Yoddha, S. Kumar, A. K. Thakur, and N. D. Jana, “Melanoma skin cancer detection using ensemble of machine learning models considering deep feature embeddings,” Procedia Comput. Sci., vol. 235, pp. 3007–3015, 2024, doi: 10.1016/j.procs.2024.04.284. [9] X. Wang, Y. Chen, and W. Zhu, “A survey on curriculum learning,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 44, no. 9, pp. 4555–4576, Sep. 2022, doi: 10.1109/TPAMI.2021.3069908. [10] B. Ozdemir and I. Pacal, “A robust deep learning framework for multiclass skin cancer classification,”
Sci. Rep., vol. 15, art. 4938, 2025, doi: 10.1038/s41598-025-89230-7. [11] N. Codella, J. Cai, M. Abedini, R. Garnavi, A. Halpern, and J. R. Smith, “Deep learning, sparse coding, and SVM for melanoma recognition in dermoscopy images,” in Proc. Int. Workshop Mach. Learn. Med. Imaging (MLMI), vol. 9352, Lecture Notes in Computer Science, L. Zhou, L. Wang, Q. Wang, and Y. Shi, Eds. Cham, Switzerland: Springer, 2015, pp. 118–126, doi: 10.1007/978-3-319-24888-2_15. [12] Y. Wei, D. Zhang, M. Gao, A. Mulati, C. Zheng, and B. Huang, “Skin cancer detection based on machine learning,” J. Knowl. Learn. Sci. Technol., vol. 3, no. 2, pp. 72–86, 2024, doi: 10.60087/jklst.vol3.n2.p86. [13] A. Demir, F. Yilmaz, and O. Kose, “Early detection of skin cancer using deep learning architectures: ResNet-101 and Inception-v3,” in Proc. Med. Technol. Congr. (TIPTEKNO), Izmir, Turkey, 2019, pp. 1–4, doi: 10.1109/TIPTEKNO47231.2019.8972045. [14] K. Aljohani and T. Turki, “Automatic classification of melanoma skin cancer with deep convolutional neural networks,” AI, vol. 3, no. 2, pp. 512–525, 2022, doi: 10.3390/ai3020029. [15] H. C. Reis, V. Turk, K. Khoshelham, and S. Kaya, “InSiNet: A deep convolutional approach to skin cancer detection and segmentation,” Med. Biol. Eng. Comput., vol. 60, no. 3, pp. 643–662, Mar. 2022, doi: 10.1007/s11517-021-02473-0. [16] S. A. H. Shah, S. T. H. Shah, R. Khaled, A. Buccoliero, S. B. H. Shah, A. Di Terlizzi, G. Di Benedetto, and M. A. Deriu, “Explainable AI-based skin cancer detection using CNN, particle swarm optimization and machine learning,” J. Imaging, vol. 10, no. 12, p. 332, 2024, doi: 10.3390/jimaging10120332. [17] I. Pacal, B. Ozdemir, J. Zeynalov, H. Gasimov, and N. Pacal, “A novel CNN-ViT-based deep learning model for early skin cancer diagnosis,” Biomed. Signal Process. Control, vol. 104, p. 107627, 2025, doi: 10.1016/j.bspc.2025.107627. [18] K. Mamar, D. Gaceb, F. Touazi, C. A. Aouchiche, Y. Bellouche, and A. Titoun, “New CNN stacking model for classification of medical imaging modalities and anatomical organs on medical images,” in Proc. Int. Conf. Intell. Data Process. Appl. Med. Imaging (IDDM), 2023, pp. 174–188. [19] N. Laribi, D. Gaceb, F. Touazi, A. Rezoug, A. Sahad, and M. O. Reggai, “Ensemble deep learning of CNN vs. Vision Transformers for brain lesion classification on MRI images,” in Proc. Int. Conf.
Intell. Data Process. Appl. Med. Imaging (IDDM), 2024, pp. 203–219. [20] M. Khaled, F. Touazi, and D. Gaceb, “Improving breast cancer diagnosis in mammograms with progressive transfer learning and ensemble deep learning,” Arab. J. Sci. Eng., vol. 50, pp. 7697–7720, 2025, doi: 10.1007/s13369-024-09428-1. [21] Z. Rahman, M. S. Hossain, M. R. Islam, M. M. Hasan, and R. A. Hridhee, “An approach for multiclass skin lesion classification based on ensemble learning,” Informat. Med. Unlocked, vol. 25, p. 100659, 2021, doi: 10.1016/j.imu.2021.100659. [22] A. Bassel, A. B. Abdulkareem, Z. A. A. Alyasseri, N. S. Sani, and H. J. Mohammed, “Automatic malignant and benign skin cancer classification using a hybrid deep learning approach,” Diagnostics, vol. 12, no. 10, p. 2472, Oct. 2022, doi: 10.3390/diagnostics12102472. [23] C.-C. Chang, Y.-Z. Li, H.-C. Wu, and M.-H. Tseng, “Melanoma detection using XGB classifier combined with feature extraction and K-means SMOTE techniques,” Diagnostics, vol. 12, no. 7, p. 1747, 2022, doi: 10.3390/diagnostics12071747. [24] M. R. Thanka, E. B. Edwin, V. Ebenezer, K. M. Sagayam, B. J. Reddy, H. Günerhan, and H. Emadifar, “A hybrid approach for melanoma classification using ensemble machine learning techniques with deep transfer learning,” Comput. Methods Programs Biomed. Update, vol. 3, p. 100103, 2023, doi: 10.1016/j.cmpbup.2023.100103. [25] P. Natha and P. RajaRajeswari, “Advancing skin cancer prediction using ensemble models,” Computers, vol. 13, no. 7, p. 157, 2024, doi: 10.3390/computers13070157. [26] S. Gamil, F. Zeng, M. Alrifaey, M. Asim, and N. Ahmad, “An eficient AdaBoost algorithm for enhancing skin cancer detection and classification,” Algorithms, vol. 17, no. 8, p. 353, 2024, doi: 10.3390/a17080353. [27] Q. Deng, J. Beltran, and D. Lee, “Assessment of segmentation impact on melanoma classification using convolutional neural networks,” J. Comput. Sci. Eng., vol. 15, no. 3, pp. 115–124, 2021, doi: 10.5626/JCSE.2021.15.3.115. [28] P. Tschandl, C. Rosendahl, and H. Kittler, “The HAM10000 dataset, a large collection of multisource dermatoscopic images of common pigmented skin lesions,” Sci. Data, vol. 5, no. 180161, 2018, doi: 10.1038/sdata.2018.161. [29] N. Codella, D. Gutman, M. E. Celebi, B. Helba, M. Marchetti, S. Dusza, A. Kalloo, K. Liopyris, N. K.
Mishra, H. Kittler, and A. Halpern, “Skin lesion analysis toward melanoma detection: A challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), hosted by the International Skin Imaging Collaboration (ISIC),” arXiv preprint arXiv:1710.05006, 2017. doi: 10.48550/arXiv.1710.05006. [30] M. Combalia, N. C. F. Codella, V. Rotemberg, B. Helba, V. Vilaplana, O. Reiter, C. Carrera, A.
Barreiro, A. C. Halpern, S. Puig, and J. Malvehy, “BCN20000: Dermoscopic lesions in the wild,” arXiv preprint arXiv:1908.02288, 2019. Available: arxiv.org/abs/1908.02288. [31] V. Rotemberg et al., “A patient-centric dataset of images and metadata for identifying melanomas using clinical context,” Sci. Data, vol. 8, p. 34, 2021, doi: 10.1038/s41597-021-00815-z.