Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contrastive meta-domain adaptation for robust skin lesion classification across clinical and acquisition conditions

Rodrigo Mota, Kelvin Cunha, Emanoel dos Santos, Fábio Papais, Francisco Filho, Thales Bezerra, Erico Medeiros, Paulo Borba, Tsang Ing Ren

TL;DR

The paper tackles degraded performance of skin lesion classifiers under acquisition and domain shifts by proposing a visual meta-domain adaptation framework. It combines a multi-transform contrastive pre-training stage to align dermoscopic and clinical representations with invariant features, and a guided-tuning domain adaptation mechanism that calibrates models across related datasets while preserving knowledge from past domains. Empirical results across HAM10000, PAD-UFES-20, and DDI show consistent accuracy and F1-score gains, reduced generalization gaps between dermoscopic and clinical images, and improved robustness under real-world imaging artifacts. The approach offers a practical pathway for deploying dermatology AI systems that remain reliable across diverse clinical workflows and devices.

Abstract

Deep learning models for dermatological image analysis remain sensitive to acquisition variability and domain-specific visual characteristics, leading to performance degradation when deployed in clinical settings. We investigate how visual artifacts and domain shifts affect deep learning-based skin lesion classification. We propose an adaptation strategy, grounded in the idea of visual meta-domains, that transfers visual representations from larger dermoscopic datasets into clinical image domains, thereby improving generalization robustness. Experiments across multiple dermatology datasets show consistent gains in classification performance and reduced gaps between dermoscopic and clinical images. These results emphasize the importance of domain-aware training for deployable systems.

Contrastive meta-domain adaptation for robust skin lesion classification across clinical and acquisition conditions

TL;DR

The paper tackles degraded performance of skin lesion classifiers under acquisition and domain shifts by proposing a visual meta-domain adaptation framework. It combines a multi-transform contrastive pre-training stage to align dermoscopic and clinical representations with invariant features, and a guided-tuning domain adaptation mechanism that calibrates models across related datasets while preserving knowledge from past domains. Empirical results across HAM10000, PAD-UFES-20, and DDI show consistent accuracy and F1-score gains, reduced generalization gaps between dermoscopic and clinical images, and improved robustness under real-world imaging artifacts. The approach offers a practical pathway for deploying dermatology AI systems that remain reliable across diverse clinical workflows and devices.

Abstract

Deep learning models for dermatological image analysis remain sensitive to acquisition variability and domain-specific visual characteristics, leading to performance degradation when deployed in clinical settings. We investigate how visual artifacts and domain shifts affect deep learning-based skin lesion classification. We propose an adaptation strategy, grounded in the idea of visual meta-domains, that transfers visual representations from larger dermoscopic datasets into clinical image domains, thereby improving generalization robustness. Experiments across multiple dermatology datasets show consistent gains in classification performance and reduced gaps between dermoscopic and clinical images. These results emphasize the importance of domain-aware training for deployable systems.
Paper Structure (14 sections, 4 equations, 5 figures, 1 table)

This paper contains 14 sections, 4 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Problem formulation
  3. Methodology
  4. Motivation
  5. Adressing task artifacts
  6. Clinically Related Artifacts
  7. Domain shift related artifacts
  8. Dataset and models validation
  9. Experiments and discussions
  10. Constrative pre-training
  11. Domain adaptation to clinical datasets
  12. Conclusion
  13. Compliance with ethical standards
  14. Acknowledgments

Figures (5)

  • Figure 1: Overview of the domain adaptation strategy. Instead of directly transferring features from a previously trained dataset (left), we introduce meta-domain adaptation that aligns the visual representations of the source domain with those of the target domain (right).
  • Figure 2: Color-based appearance transformations applied to the source domain to generate $K$ meta-domains, derived from randomly selected calibration subsets for domain adaptation.
  • Figure 3: Contrastive pre-training on HAM10000 test set. Naive training is compared with contrastive optimization to mitigate robustness degradation in visual domain features.
  • Figure 4: Validation of guided-tuning on PAD increasing the training set size.
  • Figure 5: Comparison between the guided-tuning and naive training. Naive backpropagation adapts to the current domain but exhibits forgetting of previously learned domains and poor generalization to unseen ones.