Meta-Learning-Based Few-Shot Classification Model for Accurate Multi-Class Skin Cancer Detection

In order to automate multi-class skin cancer classification, this study proposes a deep metric-based few-shot learning framework, targeting a number of limitations inherent to conventional deep learning approaches with their heavy dependence on large manuallyannotated datasets. By leveraging Siamese Networks, the proposed method learns a discriminative embedding space in which class prototypes-computed as the mean of support sample embeddings-serve as reference vectors for classifying unseen query images based on Euclidean distance. The proposed architecture was trained and evaluated on the ISIC 2020 dataset organized into seven dermoscopic skin lesion categories. Through episodic training with minimal data availability, the model generalizes exceptionally well across classes under minimal supervision. The proposed system achieved classification accuracy of 91.00% on binary (benign vs. malignant) classification tasks and 97.14% on multi-class classification task with 7 classes in input. The experimental results show that the framework holds excellent promise for such lowdata regimes, providing a scalable solution for data efficiency in clinical dermatology.