Inverse Gini indexed averaging: A multi-leveled ensemble approach for skin lesion classification using...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate skin cancer detection is crucial for effective treatment.
• Current methods face challenges in accuracy and explainability.
• Deep learning offers potential for improved skin lesion analysis.

Purpose of the Study:

• To enhance the accuracy and explainability of skin lesion detection and classification.
• To develop a novel approach using attention-integrated customized ResNet variants (CRVs) and ensemble learning (EL).
• To address the need for effective weight optimization in ensemble models for skin cancer diagnosis.

Main Methods:

• Utilized ResNet variants with channel attention, soft attention, and squeeze-excitation attention.
• Employed a multi-level ensemble learning strategy with an innovative weight optimization method, multi-leveled inverse Gini indexed averaging (ML-IGIA).
• Incorporated gradient class activation maps for model interpretability.

Main Results:

• Achieved a superior accuracy of 94.52% on the Human Against Machines 10000 dataset using the ML-IGIA approach.
• Outperformed existing methods in skin lesion classification accuracy.
• Demonstrated enhanced model transparency through interpretable heatmaps.

Conclusions:

• The proposed CRV-based ensemble model with ML-IGIA offers robust performance for skin lesion classification.
• The study addresses a research gap in weight optimization for ensemble learning.
• This approach supports timely, automated skin disease detection with high accuracy and interpretability.