CNN-Based Ensemble Architectures with Explanainable AI for Cutaneous Melanoma Identification

Malignant melanoma, a highly aggressive form of skin cancer, poses a significant global health challenge due to its rapid progression and high mortality rate if not detected on time. Early diagnosis is crucial for improving patient outcomes. The effectiveness of skin cancer detection still faces serious challenges, like visual inspection that is less accurate and time-consuming. However, deep learning-based models provide early and accurate diagnosis, serving as a supporting tool for dermatologists. Thus, this study focuses on indicating the most suitable model for skin diseases identification. Three prominent, pre-trained deep learning models, ResNet152, DenseNet201 and EfficientNet-B4, were involved in order to detect benign and malignant melanoma skin lesions. The study was performed utilizing a combined ISIC datasets gathered between 2018 and 2020 that consist of dermoscopic images. The above-mentioned deep learning algorithms were verified using accuracy, precision, recall, and F1-score metrics. Moreover, in this study the performance of skin cancer detection was enhanced utilizing soft, hard voting, and XGBoost ensemble learning methods. Combining two and three models were verified. The single models obtained accuracy at the level of 89.20%, 88.20%, and 90.40% for ResNet152, DenseNet201 and EfficientNet-B4, respectively. The soft voting ensemble, merging ResNet-152 with EfficientNet-B4 or all three models, achieved the highest absolute accuracy of 91.30%, demonstrating superior performance in melanoma diagnosis compared to individual models. Hard voting and XGBoost stated to be less effective in melanoma diagnosis. To confirm that the models were making decisions based on the significant image regions representing skin lesions, a visual explainable technique was applied. Gradient-weighted Class Activation Mapping proved the models to focus their attention to the relevant disease features. These findings underscore the potential of combining individual model strengths through ensemble learning to achieve superior diagnostic performance in melanoma detection, supporting clinicians in making more accurate and timely diagnosis.