Towards Interpretable Skin Lesion Classification with Deep Learning Models

Area of Science:

• Dermatology
• Computer Vision
• Artificial Intelligence

Background:

• Skin diseases are common globally.
• Early diagnosis is crucial for effective treatment.
• Computer vision offers potential for automated skin lesion screening.

Purpose of the Study:

• To develop an accurate and interpretable deep learning pipeline for automatic skin lesion classification.
• To improve upon existing methods for skin disease diagnosis.

Main Methods:

• Ensemble of deep learning architectures for enhanced classification accuracy.
• Integration of Generative Adversarial Networks (GANs) to augment dataset scale and diversity.
• Application of Local Interpretable Model-Agnostic Explanations (LIME) for result interpretability.

Main Results:

• Demonstrated effectiveness and robustness on a real-world skin image dataset.
• Achieved high accuracy in classifying skin lesions.
• Provided explainable classifications supporting clinical application.

Conclusions:

• The proposed deep learning pipeline offers an accurate and interpretable approach to skin lesion classification.
• The combination of ensemble models, GANs, and LIME enhances diagnostic capabilities.
• Explainability increases the potential for clinical adoption in real-world practice.