Enhancing Skin Lesion Classification Using Few-Shot Learning: Addressing Data Scarcity in Rare Lesion Diagnosis

Abstract

 Accurate skin lesion detection and identification of types of lesions is crucial.  It can also be difficult because of the scarcity of clinical data, especially for uncommon types of skin lesions. Classification between various skin conditions that may look similar visually is also a challenging task. In order to address these issues, we have proposed a skin lesion classification methodology built on a Siamese network with a fusion of attention mechanisms and few-shot learning to focus on identifying important and distinctive features from the images. Our classification framework is designed to perform well even with limited training data, allowing it to classify skin lesions into multiple classes even in challenging and complex scenarios. This methodology of multi-class classification is based on a novel approach to learning efficiently from sparse data by combining few-shot learning, attention mechanisms, and prototype-based ensemble classification. For classification, the proposed algorithm was trained on the HAM10000 image dataset, and the testing was performed on the ISIC 2018 skin lesion dataset. In this approach, we performed multi-class classification using Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbor (KNN), and XGBoost classifiers, as well as a Voting classifier. Our classification method integrates these individual classifiers into one ensemble model. The integration of a Siamese network with a Voting Classifier produced strong classification performance across all evaluation metrics. Experimental evaluation demonstrated consistent performance, achieving an accuracy of 97.42%, a precision of 97.31%, a recall of 97.55%, and an F1-score of 97.38%.These findings demonstrate the effectiveness of the hybrid approach in reliably distinguishing different categories of skin lesions.The framework is particularly suitable for applications where the availability of labeled medical data is limited, providing a robust solution that can assist clinicians in improving diagnostic consistency and supporting early identification of skin lesions.