Transformer-Based Foundation Learning for Robust and Data-Efficient Skin Disease Imaging

Area of Science:

• Artificial Intelligence in Dermatology
• Medical Image Analysis
• Deep Learning for Healthcare

Background:

• Automated dermoscopic lesion classification faces challenges with dataset bias, limited expert data, and poor generalization.
• These limitations hinder the clinical deployment of AI diagnostic systems across diverse settings and populations.

Purpose of the Study:

• To propose a transformer-based, dermatology-specific foundation model for robust dermoscopic lesion classification.
• To leverage self-supervised pretraining on unlabeled data to learn transferable visual representations.

Main Methods:

• Developed a dermatology-specific foundation model integrating large-scale self-supervised learning with a hierarchical vision transformer.
• Pretrained the model on unlabeled dermoscopic images to capture fine-grained textures and global patterns.
• Evaluated performance across ISIC 2018, HAM10000, and PH2 datasets in various settings (in-dataset, cross-dataset, limited-label).

Main Results:

• Achieved high in-dataset accuracies (94.87%-98.17%) outperforming baseline models.
• Demonstrated consistent performance gains (3.5-5.8%) in cross-dataset transfer, indicating improved robustness to domain shift.
• Attained performance comparable to fully supervised methods with only 10% labeled data, highlighting strong data efficiency.

Conclusions:

• Dermatology-specific foundation learning provides a practical solution for robust dermoscopic lesion classification.
• The proposed model addresses realistic clinical constraints, including limited labeled data and domain variability.
• This approach paves the way for more reliable AI-powered diagnostic tools in clinical dermatology.