Addressing Data Scarcity in Dermatology: A Systematic Literature Review of Few-Shot Learning from Metric Learning to Generative Models

Abstract

Background: Dermoscopy, a non-invasive imaging technique magnifying subsurface skin structures, has become the gold standard for early skin cancer detection, reducing diagnostic errors by up to 49% compared to naked-eye examination. While Deep Learning (DL) models now achieve dermatologist-level accuracy on common malignancies like melanoma when trained on large-scale datasets (e.g., ISIC archive with tens of thousands of images), clinical dermatology encompasses over 2,000 distinct conditions following a “long-tail” distribution. This creates a critical AI divide: rare, neglected, and emerging tropical diseases lack sufficient labeled data for standard supervised learning, rendering conventional DL approaches fundamentally ill-suited and leaving thousands of rare variants diagnostically underserved. Problem: The scarcity of annotated dermoscopic images for rare skin diseases poses a severe barrier to robust Medical AI deployment, as traditional deep learning models require thousands of examples per class and catastrophically overfit or exhibit severe bias in low-data regimes. This discrepancy creates incomplete clinical safety and limited specialist utility, particularly in resource-constrained settings where diagnostic expertise is scarce. Objective: This Systematic Literature Review (SLR) investigates how Few-Shot Learning (FSL) and Meta-Learning methods have been designed, validated, and applied to bridge the data scarcity gap in dermatological diagnosis between 2020–2025, analyzing their robustness, generalization capabilities, and clinical readiness for equitable healthcare delivery across the full disease spectrum. Method: Following PRISMA 2020 guidelines, we systematically searched IEEE Xplore, Scopus, ScienceDirect, and PubMed, analyzing 16 primary studies using strict PICOC eligibility criteria and a custom 10-item quality assessment framework adapted from QUADAS-2. Results: We identify a clear technological evolution from early Metric-based methods (2020–2021) to advanced Generative Foundation Models (2024–2025), with key findings highlighting the effectiveness of Generative Adversarial Networks (GANs) for feature-level hallucination and Vision Transformer backbones for Domain Generalization. However, significant barriers persist: only 19% of studies provide publicly available code, exposing a reproducibility crisis, and frequent data leakage from improper patient-level splitting undermines reported performance. Future Direction: While generative approaches and parameter-efficient fine-tuning offer promising pathways for rare disease diagnosis, achieving clinical deployment requires prioritizing external validation across diverse populations, rigorous patient-level data splits, standardized evaluation protocols, and open science practices to ensure reproducibility, fairness, and real-world clinical utility.