Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Meta-Transfer Derm-Diagnosis: Exploring Few-Shot Learning and Transfer Learning for Skin Disease Classification in Long-Tail Distribution

Zeynep Özdemir, Hacer Yalim Keles, Ömer Özgür Tanrıöver

TL;DR

The paper tackles the challenge of classifying rare skin diseases under long-tailed data with limited labels. It compares three learning paradigms—episodic learning, supervised transfer learning, and contrastive self-supervised pretraining—across ISIC2018, Derm7pt, and SD-198 within a few-shot framework. The results show that traditional transfer learning using MobileNetV2 and Vision Transformer architectures consistently outperforms episodic and self-supervised methods as training data increases, and that batch-level augmentations like MixUp, CutMix, and ResizeMix further boost performance, achieving state-of-the-art on SD-198 and Derm7pt and highly competitive results on ISIC2018. The findings offer practical guidance for deploying dermatology classifiers under data scarcity, and the authors plan to release source code publicly.

Abstract

Building accurate models for rare skin diseases remains challenging due to the lack of sufficient labeled data and the inherently long-tailed distribution of available samples. These issues are further complicated by inconsistencies in how datasets are collected and their varying objectives. To address these challenges, we compare three learning strategies: episodic learning, supervised transfer learning, and contrastive self-supervised pretraining, within a few-shot learning framework. We evaluate five training setups on three benchmark datasets: ISIC2018, Derm7pt, and SD-198. Our findings show that traditional transfer learning approaches, particularly those based on MobileNetV2 and Vision Transformer (ViT) architectures, consistently outperform episodic and self-supervised methods as the number of training examples increases. When combined with batch-level data augmentation techniques such as MixUp, CutMix, and ResizeMix, these models achieve state-of-the-art performance on the SD-198 and Derm7pt datasets, and deliver highly competitive results on ISIC2018. All the source codes related to this work will be made publicly available soon at the provided URL.

Meta-Transfer Derm-Diagnosis: Exploring Few-Shot Learning and Transfer Learning for Skin Disease Classification in Long-Tail Distribution

TL;DR

The paper tackles the challenge of classifying rare skin diseases under long-tailed data with limited labels. It compares three learning paradigms—episodic learning, supervised transfer learning, and contrastive self-supervised pretraining—across ISIC2018, Derm7pt, and SD-198 within a few-shot framework. The results show that traditional transfer learning using MobileNetV2 and Vision Transformer architectures consistently outperforms episodic and self-supervised methods as training data increases, and that batch-level augmentations like MixUp, CutMix, and ResizeMix further boost performance, achieving state-of-the-art on SD-198 and Derm7pt and highly competitive results on ISIC2018. The findings offer practical guidance for deploying dermatology classifiers under data scarcity, and the authors plan to release source code publicly.

Abstract

Building accurate models for rare skin diseases remains challenging due to the lack of sufficient labeled data and the inherently long-tailed distribution of available samples. These issues are further complicated by inconsistencies in how datasets are collected and their varying objectives. To address these challenges, we compare three learning strategies: episodic learning, supervised transfer learning, and contrastive self-supervised pretraining, within a few-shot learning framework. We evaluate five training setups on three benchmark datasets: ISIC2018, Derm7pt, and SD-198. Our findings show that traditional transfer learning approaches, particularly those based on MobileNetV2 and Vision Transformer (ViT) architectures, consistently outperform episodic and self-supervised methods as the number of training examples increases. When combined with batch-level data augmentation techniques such as MixUp, CutMix, and ResizeMix, these models achieve state-of-the-art performance on the SD-198 and Derm7pt datasets, and deliver highly competitive results on ISIC2018. All the source codes related to this work will be made publicly available soon at the provided URL.
Paper Structure (1 section, 2 figures)

This paper contains 1 section, 2 figures.

Table of Contents

  1. Introduction

Figures (2)

  • Figure 1: Skin lesion classification framework: (a) benchmark datasets, (b) transfer learning strategies.
  • Figure 2: The figure shows the distributions of the datasets SD-198sd198dataset, Derm7ptderm7ptdataset, and ISIC2018isic2018dataset, highlighting their long-tailed nature with some classes having very few instances. Base classes (common diseases) are marked as train (green) and validation (yellow), while novel classes (rare diseases) are labeled as test (red). In the Derm7pt dataset, classes with very few examples are shown as deleted (grey) and excluded from use.