Addressing Class Imbalance in Skin Lesion Classification: Few-Shot Learning vs Traditional CNNs

Identifier:  TFM:2533

Handle: https://hdl.handle.net/20.500.11797/TFM2533

Authors:  Iturrate Reyzabal, Ander De

Abstract:
This Master’s Thesis investigates machine learning models for automated skin lesion classification using dermoscopic images, focusing on class imbalance and limited data availability. Traditional CNNs, advanced architectures like Inception v3 and EfficientNet-B3, and Few-Shot Learning (FSL) based on Prototypical Networks were evaluated. Fine-tuning pretrained models improved performance, especially for minority classes such as dermatofibroma and vascular lesions. EfficientNet-B3 achieved 84.7% accuracy with a macro F1-score of 76.6%. FSL effectively handled scarce data, enabling robust classification with few support samples. The study emphasizes integrating transfer and meta-learning to build generalizable, reliable models for AI-assisted dermatological diagnostics.