Attention-Guided Ensemble Deep Learning Framework for Automated Skin Cancer Classification

Abstract

Skin cancer is one of the most common malignancies in the world and early and accurate dermoscopic diagnosis is crucial for better survival outcomes of patients. There are various limitations in current single model convolutional and transformer models, such as limited ability to capture local texture, multi-scale morphology, and global contextual information, as well as high intra-class visual similarity and extreme class imbalance. To overcome these problems, this paper proposes an Attention-Guided Ensemble Deep Learning (AGEDL) framework which integrates the EfficientNet-B3, InceptionV3 and Swin Transformer to simultaneously learn complementary dermoscopic representations from the seven-class ISIC 2018 dataset. The Squeeze-and-Excitation (SE) attention blocks dynamically modulate feature responses in each channel, which reduces background noise and enhances discriminative features of lesions. The Lion optimizer offers stable and efficient training in both base training and fine-tuning stages. An XGBoost stacking metalearner is used to combine all backbone networks, and hyperparameters of the XGBoost are optimized by the Dhole Optimizer. The proposed AGEDL achieved 98.02% accuracy, 97.09% precision, 97.05% recall and 94.07% F1 score with 95% confidence interval of 97.41%–98.59%, surpassing the state-of-the-art CNN, transformer-based and hybrid ensemble baselines.