XAI-MedNet: A Next-Generation Explainable AI Framework for Contrast-Enhanced Skin Lesion Classification via Entropy-Controlled Optimization

Explainable Artificial Intelligence (XAI) has become a critical requirement in medical image analysis, where transparency and interpretability are essential for clinical trust and decision support. Melanoma is recognized as one of the most deadly types of skin cancer, with its occurrence exhibiting an increasing pattern in recent times. However, detecting this cancer in its initial stages greatly increases patients’ chances of long-term survival. Various computer-based techniques have recently been proposed to diagnose skin lesions at their early stages. Even though the machine learning community has achieved a certain degree of success, there is still an unresolved research challenge regarding high error margins and the limited interpretability of automated systems. This study focuses on addressing both segmentation and classification tasks, with particular emphasis on two key concepts: (1) improving image quality to maximize distinguishability between foreground and background regions, thereby enhancing visual interpretability and segmentation accuracy and (2) eliminating redundant and cluttered feature information to generate the most discriminative and compact feature representations. The input images are initially processed using a novel metaheuristic contrast-stretching method to estimate image-specific key parameters, thereby enhancing lesion boundary clarity in a clinically interpretable manner. Following this, the improved images are fed into selected pre-trained deep models, including DenseNet-201, Inception-ResNet v2, and NASNet-Mobile. The extracted features from all pre-trained models are fused to produce resultant vectors, which are then refined using a bio-inspired feature selection method, termed entropy-controlled whale optimization, to retain only the most informative attributes. The selected discriminative feature set is subsequently classified using multiple classifiers. The results indicate that the proposed framework achieves superior performance compared to existing methods in terms of accuracy, sensitivity, specificity, and F1-score. Additionally, it facilitates a more explainable, transparent, and structured diagnostic pipeline appropriate for medical applications.