An Explainable Privacy Preserving Multimodal Ensemble Framework For Skin Lesion Classification

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer poses a significant threat, necessitating early and accurate diagnosis.
• Traditional AI diagnostic methods struggle with privacy, interpretability, and data imbalance.
• Class imbalance in multi-class skin lesion datasets hinders diagnostic accuracy.

Purpose of the Study:

• To develop a privacy-aware, explainable multimodal skin lesion classification model.
• To address challenges in traditional AI-based skin cancer diagnosis.
• To improve the accuracy and transparency of AI-driven dermatological diagnostics.

Main Methods:

• Utilized a class-balancing technique to address data imbalance in the HAM10000 dataset.
• Employed EfficientNet B4, DenseNet201, and MobileNetv2 for deep feature extraction.
• Combined deep features with clinical metadata to create a multimodal feature space for XGBoost, LightGBM, and Deep Neural Classifier (DNC) training.
• Implemented a stacked ensemble strategy to integrate model outputs for enhanced classification accuracy.
• Applied model interpretability techniques for feature-level explanations.

Main Results:

• Individual models achieved classification accuracies of 92% (XGBoost), 90% (LightGBM), and 94% (DNC).
• The stacked ensemble model achieved a final classification accuracy of 96%.
• The framework demonstrated practical efficiency and clinical relevance in skin lesion classification.
• Feature-level explanations enhanced model transparency.

Conclusions:

• The proposed privacy-aware, explainable multimodal model effectively classifies skin lesions with high accuracy.
• Ensemble modeling and multimodal feature integration significantly improve diagnostic performance.
• The approach offers a practical, transparent, and efficient solution for AI-assisted dermatological diagnosis.
• This framework holds promise for improving early skin cancer detection and patient prognosis.