Multi-Model Attentional Fusion Ensemble for Accurate Skin Cancer Classification

Skin cancer, with its rising global prevalence, remains a crucial healthcare challenge, necessitating efficient and early detection for better patient outcomes. While deep convolutional neural networks have advanced image classification, current models struggle with diverse lesion types, variable image quality, and dataset imbalances. Artifacts like hair can further obscure important features. This research addresses the problem and introduces a novel deep learning approach for accurate skin cancer classification by combining ResNet50V2, MobileNetV2, and EfficientNetV2 models. Our proposed architecture leverages the unique feature extraction capabilities of these models. It incorporates an attention mechanism to dynamically emphasize relevant features, enhancing focus, and promoting synergistic interactions among diverse feature sets. As such, our ensemble architectural approach outperforms other state-of-the-art models with high precision, recall, and F1-score metrics. Notably, the model demonstrates significant precision for Dermatofibroma (92% with 96% recall) and Vascular lesions (99% with 99% recall), highlighting its robustness across varied lesion types. Additionally, comprehensive image preprocessing techniques, including image resampling, black-hat filtering, thresholding, morphological closing, inpainting, and overall hair artifact removal, ensure the dataset's quality and reliability. Validated through statistical significance testing and prototyped with an mHealth solution, this research heralds a significant stride in skin cancer diagnosis with the potential of attention-enhanced ensemble architectures.