A Multi-level ensemble approach for skin lesion classification using Customized Transfer Learning with Triple...

Area of Science:

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

Background:

• Skin lesions, including skin cancer, require early detection for effective treatment.
• Current automated methods struggle with accurate lesion identification and region localization.
• Deep learning models often lack interpretability, hindering clinical trust and adoption.

Purpose of the Study:

• To develop an accurate and interpretable deep learning framework for skin lesion detection.
• To introduce a novel method for optimal model weight aggregation in ensemble learning.
• To address the challenge of identifying specific regions responsible for model predictions.

Main Methods:

• A Convolutional Neural Network (CNN) framework utilizing Customized Transfer Learning (CTL) and Triple Attention (TA) modules.
• Development of Multi-Level Information Gain Proportioned Averaging (ML-IGPA) for optimal ensemble weighting.
• Application of Gradient Class Activation Map (GradCAM) for model interpretability and region identification.

Main Results:

• The proposed ML-IGPA approach achieved 94.93% accuracy on the HAM1000 dataset.
• The method surpassed existing state-of-the-art techniques in skin lesion classification.
• GradCAM successfully visualized the regions influencing model predictions, enhancing explainability.

Conclusions:

• The developed deep learning model significantly improves skin lesion detection accuracy and interpretability.
• The ML-IGPA technique offers a robust solution for optimizing ensemble model performance.
• This approach facilitates earlier and more reliable diagnosis of skin lesions, with potential to reduce healthcare costs and improve patient outcomes.