Adversarial Training Based Domain Adaptation of Skin Cancer Images.

Area of Science:

• Medical image analysis
• Artificial intelligence in dermatology

Background:

• Skin lesion datasets often suffer from class imbalance, impacting deep learning model performance.
• Generative Adversarial Networks (GANs) can create synthetic data but may introduce bias and domain shift.
• Variations in imaging instruments and resolution further exacerbate domain shifts in skin lesion datasets.

Purpose of the Study:

• To develop and evaluate a domain adaptation methodology for mitigating bias and domain shift in skin lesion datasets.
• To improve the performance of deep learning models on imbalanced and shifted skin lesion data.

Main Methods:

• Implementation of a domain adaptation algorithm to address dataset bias and domain shift.
• Conducted six experiments utilizing two distinct domain adaptation architectures.
• Employed a domain adversarial neural network with gradient reversal layers and VGG13 as a feature extractor.

Main Results:

• The domain adversarial neural network architecture demonstrated superior performance.
• Achieved the highest accuracy of 0.7567 and F1 score of 0.75.
• Resulted in an 18.47% improvement in accuracy compared to the baseline model.

Conclusions:

• Domain adaptation algorithms are effective in mitigating bias and domain shift in skin lesion datasets.
• The proposed methodology enhances the robustness and accuracy of deep learning models for skin lesion analysis.
• The domain adversarial neural network with VGG13 shows significant potential for clinical application.