Towards Accurate and Explainable Skin Cancer Identification: A Comparative Analysis of Machine Learning, Deep Learning, Vision Transformers and Hybrid XAI-Based Methods

This study presents a comprehensive framework for comparing machine-learning and deep-learning methods for automated skin lesion identification. Conventional classifiers such as SVM and Random Forest provide useful baselines, but deep neural networks can learn richer representations directly from dermoscopic images. Transfer learning further improves the practicality of deep models when medical datasets are comparatively limited. The proposed EHSC-XAI framework extends conventional classification by combining preprocessing, class-aware augmentation, multi-scale feature representation, attention-based fusion, metadata integration, ensemble prediction, confidence estimation, and explainable AI. The objective is not simply to maximize an accuracy value but to develop a system that can be evaluated from several perspectives. A particularly important aspect of the proposed methodology is the separation between prediction performance and explanation quality. A model may classify an image correctly for the wrong reason. Conversely, an apparently modest classification improvement may be more valuable if the model becomes better calibrated and more consistent across lesion categories. The study therefore recommends reporting macro-F1, balanced accuracy, class-wise sensitivity, specificity, ROC-AUC, PR-AUC, calibration, computational cost, and XAI behavior alongside conventional accuracy. The use of HAM10000 as the principal benchmark and ISIC data for external assessment provides a practical experimental foundation. HAM10000 was developed specifically to provide a large, diverse collection of dermoscopic images for machine-learning research, while ISIC has provided standardized challenge environments for lesion analysis and classification. The final effectiveness of EHSC-XAI, however, must be demonstrated through actual experiments. Numerical results should therefore be obtained from controlled training and testing rather than inserted from existing publications. The proposed ablation study will make it possible to determine whether preprocessing, augmentation, attention, metadata, and ensemble learning genuinely contribute to the improvement. Ultimately, the most useful skin cancer identification system is unlikely to be the model with the highest single-dataset accuracy. A stronger candidate is a system that maintains performance across datasets, recognizes uncertainty, performs reasonably across minority classes, highlights relevant image regions, and communicates its limitations clearly. Such characteristics can provide a stronger foundation for future clinical validation and responsible deployment of artificial intelligence in dermatology.

[1] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, “Dermatologist-level classification of skin cancer with deep neural networks,” Nature, vol. 542, pp. 115–118, 2017. [2] H. A. Haenssle et al., “Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists,” Annals of Oncology, vol. 29, no. 8, pp. 1836–1842, 2018. [3] P. Tschandl, C. Rosendahl, and H. Kittler, “The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented lesions,” Scientific Data, vol. 5, Article 180161, 2018. [4] N. C. F. Codella et al., “Skin lesion analysis toward melanoma detection: A challenge at the 2017 International Symposium on Biomedical Imaging,” in Proc. IEEE ISBI, 2017. [5] N. C. F. Codella et al., “Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the International Skin Imaging Collaboration,” 2019. [6] N. Gessert, M. Nielsen, M. Shaikh, R. Werner, and A. Schlaefer, “Skin lesion classification using ensembles of multi-resolution EfficientNets with meta data,” MethodsX, 2020. [7] M. Tan and Q. V. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proc. ICML, 2019. [8] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE CVPR, 2016. [9] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proc. IEEE CVPR, 2017. [10] C. Szegedy et al., “Rethinking the inception architecture for computer vision,” in Proc. IEEE CVPR, 2016.