Skin Lesion Classification Breakthrough: Leveraging Independent-Serial–Parallel–Stacking Ensemble Architecture With Reciprocal Cross-Entropy Averaging

Early detection of skin lesions, such as malignancies and precancerous anomalies, is vital for timely intervention, yet hindered by class imbalance in datasets and the absence of augmentation protocols. Contemporary approaches further struggle with suboptimal feature localization due to model dominance dispersion and limited adaptability of pretrained architectures to specialized diagnostic tasks. This work introduces Independent-Serial-Parallel-Stacking Ensemble (ISPSE), a novel transfer learning framework integrating Inception and MobileNet architectures via three innovative stacking paradigms: Independent-Serial-Parallel Stacking Ensemble with Triple-Attention (TA). The TA mechanism synergizes Self-Attention, Channel Attention, and Squeeze-Excitation Attention to amplify discriminative feature extraction while mitigating spatial redundancy. Complementing this, we propose Reciprocal Cross-Entropy Averaging (RCEA), extended into Multi-Layer RCEA (ML-RCEA) to dynamically optimize prediction weighting across hierarchical layers—addressing a critical gap in ensemble-based diagnostic systems. Evaluated on the ISIC2018 dataset, ISPSE by ML-RCEA achieves state-of-the-art accuracy (90.96%), surpassing contemporary benchmarks by a substantial margin. Gradient Class Activation Maps (Grad-CAM) further validate the framework’s clinical relevance by visually delineating diagnostically salient regions, enhancing interpretability for practitioners. Beyond accuracy, our framework’s modular design and adaptive augmentation strategies mitigate challenges such as dataset imbalance, limited accessibility, and computational costs. This study advances automated dermatological diagnostics by unifying architectural innovation, ensemble optimization, and explainable AI, offering a paradigm shift toward deployable, high-confidence systems. The methodology and findings provide actionable insights for real-world clinical integration, bridging the gap between theoretical machine learning and life-saving medical applications.