Enhancing Skin Lesion Classification Generalization with Active Domain Adaptation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Dermatology

Background:

• Deep learning models for skin lesion classification often struggle with generalization across different datasets.
• Domain shift, variations in data distribution between datasets, is a major challenge.

Purpose of the Study:

• To develop a method that improves the generalization capability of skin lesion classification models.
• To address the challenge of domain shift in medical image analysis.

Main Methods:

• Combining self-supervised learning (SSL) with active domain adaptation (ADA).
• Pre-training on natural images, retraining on skin lesion datasets, fine-tuning on source data, and applying ADA on target data.
• Evaluating the approach on ten diverse skin lesion datasets using five ADA methods.

Main Results:

• The proposed approach demonstrated improved generalization performance for skin lesion classification.
• Effectiveness was shown across various domain shifts and dataset configurations.
• The combination of SSL and ADA proved beneficial in enhancing model robustness.

Conclusions:

• The integrated SSL and ADA method offers a promising solution for improving skin lesion classification model generalization.
• This approach can facilitate the wider clinical implementation of AI in dermatology and other medical imaging fields.
• Addressing domain shift is crucial for reliable AI deployment in healthcare.