Explainability agreement between dermatologists and five visual explanations techniques in deep neural networks for...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Dermatology

Background:

• Deep convolutional neural networks (CNNs) show promise for skin lesion analysis and early skin cancer detection.
• Clinical implementation requires model explainability for specialist trust and understanding.
• Explainability techniques highlight patterns relevant to model predictions.

Purpose of the Study:

• To evaluate five explainability techniques (Grad-CAM, Grad-CAM++, Score-CAM, Eigen-CAM, LIME) for analyzing skin lesion images.
• To assess the agreement between AI-identified features and clinical criteria (asymmetry, border irregularity, color heterogeneity) in melanoma classification.
• To determine dermatologist preferences for explainability methods.

Main Methods:

• 100 melanoma images were analyzed using five explainability techniques.
• Two dermatologists scored visual explanation maps against clinical images using a semi-quantitative scale.
• Agreement rates and technique rankings were compared.

Main Results:

• Grad-CAM demonstrated the highest agreement rate (93.6%) with clinical criteria.
• LIME, Grad-CAM++, Eigen-CAM, and Score-CAM showed varying agreement rates.
• Dermatologists favored Grad-CAM and Grad-CAM++ most highly.

Conclusions:

• Explainability methods, particularly Grad-CAM, show significant agreement with clinical features used by dermatologists for melanoma diagnosis.
• Human evaluation of explainability is vital for assessing clinical applicability.
• These findings support the integration of explainable AI in dermatological diagnostics.