Developing a trustworthy and explainable framework for classifying skin lesions through transfer learning and...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate detection of skin lesions, particularly melanoma, remains a significant challenge in medical imaging.
• Limited availability of reliably labeled datasets hinders diagnostic precision.
• Existing methods often struggle with the visual similarities between benign and malignant lesions.

Purpose of the Study:

• To develop a deep learning framework for precise skin lesion detection and classification.
• To enhance clinical accuracy and interpretability in AI-driven dermatological diagnosis.
• To create a reliable tool for integrating artificial intelligence into clinical dermatology practice.

Main Methods:

• A U-NET segmentation model was employed for preprocessing, including hair removal and lesion segmentation.
• A modified EfficientNet-B4 network with a CBAM module was utilized for feature extraction.
• The model was integrated with ResNet-50 and predictions aggregated via soft voting.
• Explainable AI techniques (SHAP and LIME) were used for model interpretation.

Main Results:

• The model achieved high performance metrics: 98.95% accuracy, 98.7% balanced accuracy, and 99.6% sensitivity for melanoma detection.
• The framework demonstrated strong generalization capabilities across different datasets (ISIC-2019, PH2).
• Interpretability methods provided visual insights into the decision-making process of the AI model.

Conclusions:

• The developed deep learning framework offers a significant advancement in the automated detection of skin lesions.
• The combination of high accuracy, interpretability, and generalization makes the model a promising tool for clinical dermatology.
• This AI framework represents a realistic step towards the safe and effective integration of AI in dermatological practice.