Skin cancer represents a significant health issue marked by the irregular proliferation of skin cells. It is frequently associated with exposure to ultraviolet (UV) radiation. Traditional ways of diagnosing depend on visual inspection and biopsy, which are subjective and take a lot of time and money. Automating medical image analysis using artificial intelligence (AI) promises to enhance diagnostic accuracy and efficiency. Medical imaging encounters the issue of inadequate data. Few-shot learning addresses the scarcity of labeled medical data by enabling effective knowledge transfer from minimal samples. Typically, few-shot classification extracts features only from a single layer of a convolutional neural network (CNN). It may limit the model's capacity to generalize effectively across diverse tasks. This paper uses the SetFeat method with ResNet-12 to classify skin cancer in a few shots. SetFeat extracts feature sets from multiple layers of convolutional neural networks. It also incorporates shallow self-attention mechanisms to improve feature representation and discriminative capability. The method is assessed using the HAM10000 dataset. The model attains an accuracy of 63.34% for a 2-way 1-shot with 15 queries, 73.90% for a 2-way 5-shot with 15 queries, and 85.23% for a 2-way 10-shot with 15 queries.