MultiExCam: A multi approach and explainable artificial intelligence architecture for skin lesion classification.

Area of Science:

• Artificial Intelligence in Dermatology
• Medical Image Analysis
• Computational Pathology

Background:

• Cutaneous melanoma is a lethal skin cancer, but early diagnosis significantly improves survival rates.
• Existing AI solutions for skin lesion diagnosis often use machine learning and deep learning in isolation.
• There is a need for integrated AI approaches that combine multiple data sources and provide interpretable results.

Purpose of the Study:

• To introduce MultiExCam, a novel, explainable, multi-approach AI architecture for skin cancer detection.
• To integrate heterogeneous data sources including dermatoscopic images and extracted features.
• To develop an AI system that combines machine learning and deep learning for enhanced diagnostic performance and interpretability.

Main Methods:

• MultiExCam utilizes a hybrid architecture integrating deep learning (CNN) for feature extraction and initial classification with machine learning models.
• It combines deep learning features with hand-crafted statistical features for training ensemble models.
• An advanced ensemble model with gating and attention mechanisms provides the final classification, enhanced by GradCAM and SHAP for explainability.

Main Results:

• MultiExCam achieved high performance across diverse datasets with AUC scores up to 98% and F1-scores up to 94%.
• The hybrid ensemble approach outperformed baseline deep learning models by 1-3%.
• Explainability analysis identified clinically relevant patterns linked to diagnostic criteria like asymmetry and irregular borders.

Conclusions:

• MultiExCam sets a new standard for AI-assisted dermatological diagnosis by integrating deep learning, machine learning, and explainability.
• The AI's ability to provide accurate, interpretable predictions addresses key requirements for clinical adoption.
• This architecture offers a strong foundation for AI-driven clinical decision support systems in melanoma detection.