Comprehensive Summary
The study presents an optimized deep learning learning model for automatic classification of benign nevi and malignant melanoma from dermoscopic images, addressing challenges of image noise, data imbalance, and poor generalization. Using the ISIC-2019 dataset, researchers developed a modified Inception-ResNet-V2 architecture incorporating median filtering for artifact removal and Synthetic Minority Oversampling Technique (SMOTE for class balancing. The model’s performance was compared with ResNet-50, EfficientNet-B0, and Inception-V3 across multiple optimization strategies (Adam, Nadam, and AdaMax) using five-fold cross-validation. The optimized model achieved the best overall performance with the AdaMax optimizer, achieving 97.65% accuracy, 96.67% sensitivity, and 98.92% specificity, outperforming all benchmark models.
Outcomes and Implications
The optimized Inception-ResNet-V2 framework demonstrates substantial potential for non-invasive, early detection of melanoma, offering diagnostic accuracy comparable to that of dermatologists. Its robust preprocessing and data augmentation pipeline improve adaptability to real-world clinical settings, including integration with smartphone-based diagnostic systems for remote screening. By effectively managing noise and class imbalance, the model enhances diagnostic confidence, potentially reducing the need for biopsies and supporting dermatologists in clinical decision-making. Future extensions include multi-class classification for broader lesion categories, cross-dataset validation across diverse skin tones, and the inclusion of Explainable AI (XAI) tools (e.g., SHAP, LIME) to strengthen interpretability and clinical trust.