Meta-Learned Ensemble of CNN and Vision Transformers for Imbalanced Skin Cancer Classification

One of the most serious health issues facing the world today is skin cancer, such as melanoma or basal cell carcinoma and squamous cell carcinoma, because of the mortality rates when the disease is not promptly diagnosed. Deep-learning-based automatic diagnostic systems may greatly improve the early diagnosis and patient outcomes. The paper suggests a new approach based on the idea to combine a modified Convolutional Neural Network (CNN) with Vision Transformer (ViT) and EfficientNetB3 into a transformer-based ensemble detector of skin cancer. Based on the 2018 version of the ISIC dataset, the method uses advanced preprocessing and attention-based feature extraction and an ensemble technique based on gradient boosting to obtain high classification accuracy. The model tells a classification accuracy of 98.4% by resolving the issue of class imbalance in adaptive data augmentation and training optimization with cosine annealing.