Methodology for Deep Ensemble Learning and Explainable AI in Imbalanced Melanoma Classification

Automated analysis of dermoscopic images has the potential to support the early identification of melanoma, but reliable classification remains challenging because of class imbalance, variations in image quality, limited labelled data, and the difficulty of interpreting deep learning decisions. This study presents a methodological framework for developing an ensemble-based melanoma classification system with an emphasis on systematic model selection, prediction fusion, and explainability. The framework uses dermoscopic images from the ISIC 2019 and ISIC 2020 datasets and follows a sequence of data. preparation, image preprocessing, transfer-learning model development, candidate model evaluation, ensemble construction, and explanation generation. Class imbalance is addressed through majority-class downsampling, while image enhancement, centre cropping, and normalization are applied before model training. Five pretrained convolutional neural networks—VGG-19, ResNet-50, ResNet-101, DenseNet-121, and InceptionV3—are considered as candidate learners. Based on their classification performance, suitable models are selected for ensemble construction. Several prediction-fusion strategies are then examined, including hard voting, probability averaging, maximum-rule fusion, and weighted probability averaging. The weighting mechanism considers multiple evaluation measures rather than relying only on accuracy. Finally, SHAP is incorporated to examine the image regions contributing to individual predictions. This methodology provides a structured basis for evaluating whether combining complementary deep learning models can improve melanoma classification while making model decisions more interpretable