Few-shot learning for skin lesion classification: A prototypical networks approach

Abstract

Prototypical networks (PN) have emerged as one of multiple effective approaches for few-shot learning (FSL), even in medical image classification. This study focuses on implementing a PN for skin lesion classification to assess its performance, generalizability, and robustness when applied across 11 dermoscopic image domains. Unlike conventional FSL scenarios, where the performance is evaluated for unseen classes in the test set, our analysis extends this to evaluate PNs on a complete hold-out dataset with the same classes from a different domain. Differences in a patient’s age, lesion localization, or image acquisition systems variations mimic real-world cross-domain conditions in a clinic. Given the scarcity of medical datasets, this assessment is crucial for potentially translating such systems into real-world clinical settings to support physicians with the diagnosis. Our primary focus is two-fold: investigating whether a PN performs on par with a baseline classifier, even using only a limited number of reference samples from the hold-out test set (in-domain) and whether a PN can generalize to the same classes of unseen domains (cross-domain). Our analysis uncovers that a PN can perform on par with the baseline classifier in an in-domain setting, even with only a few support samples. However, in cross-domain scenarios, a PN exhibits improved performance only on specific domains, while others demonstrate similar or even decreased performance when confronted with a smaller number of images. Our findings contribute to comprehending potential opportunities and limitations of FSL in dermatological practice.