DermAI 1.0: A Robust, Generalized, and Novel Attention-Enabled Ensemble-Based Transfer Learning Paradigm for...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate skin lesion classification is vital for early detection of malignant conditions.
• Standalone transfer learning (TL) models show limitations in achieving optimal diagnostic performance.

Purpose of the Study:

• To develop a novel, generalized deep learning technique for enhanced skin lesion classification.
• To improve diagnostic accuracy by integrating attention mechanisms and ensemble methods.

Main Methods:

• Utilized seven pre-trained transfer learning (TL) models for feature extraction.
• Developed six ensemble-based deep learning (EBDL) models using stacking, softmax voting, and weighted averaging.
• Created seven attention-enabled transfer learning (aeTL) models and three attention-enabled ensemble-based deep learning (aeEBDL) models.

Main Results:

• Ensemble-based methods improved TL model accuracy from 95.30% to 99.52%.
• Attention-enabled transfer learning (aeTL) models showed a 3.01% accuracy improvement over TL models.
• Attention-enabled ensemble-based deep learning (aeEBDL) models achieved the highest accuracy, outperforming aeTL models by 1.29%.

Conclusions:

• The proposed attention-enabled ensemble-based deep learning approach is highly effective and generalized for skin lesion classification.
• The method demonstrates significant potential for enhancing diagnostic accuracy in dermatology.
• Statistical validation confirmed the reliability and effectiveness of the aeEBDL models.