Skin Cancer Detection Using Transfer Learning and Deep Attention Mechanisms

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Dermatology

Background:

• Accurate skin cancer diagnosis is vital for survival but challenging due to similar lesion pigmentation.
• Deep learning and transfer learning show promise for skin cancer image analysis.
• Attention mechanisms enhance deep learning accuracy in medical image classification.

Purpose of the Study:

• To investigate the impact of attention mechanisms (AMs) on the Xception transfer learning model for binary classification of skin lesions.
• To evaluate the performance of self-attention, hard attention, and soft attention integrated with the Xception model.
• To compare the enhanced Xception model's performance against the standard Xception model for skin cancer detection.

Main Methods:

• Four experiments were conducted using the HAM10000 dataset.
• Three models integrated self-attention, hard attention, and soft attention mechanisms with the Xception architecture.
• A baseline model used the standard Xception without attention mechanisms for comparison.

Main Results:

• The standard Xception model achieved 91.05% accuracy.
• Integrating attention mechanisms improved accuracy: self-attention (94.11%), soft attention (93.29%), and hard attention (92.97%).
• The proposed models demonstrated superior recall metrics compared to previous studies.

Conclusions:

• Attention mechanisms enhance the performance of the Xception model for skin lesion classification.
• These findings suggest potential for earlier diagnosis and improved treatment outcomes in medical imaging.
• Further research into AMs can advance AI-driven dermatological diagnostics.