Few-shot skin lesion classification with Adaptive Multi-Scale Convolutional Attention Network

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Computer-aided diagnosis of skin lesions faces challenges like scale variation, blurred boundaries, and limited data.
• Existing deep learning models struggle with the unique characteristics of skin lesions due to fixed receptive fields or generic attention mechanisms.

Purpose of the Study:

• To develop an Adaptive Multi-scale Convolutional Attention Network (AMCANet) for accurate and robust skin lesion classification with sparse data.
• To improve the adaptability of deep learning models to diverse skin lesion characteristics.

Main Methods:

• Proposed AMCANet with three core modules: adaptive multi-scale convolution, hierarchical channel attention, and skin spatial attention.
• Adaptive multi-scale convolution dynamically adjusts receptive fields for varying lesion sizes.
• Hierarchical channel attention integrates multi-level semantic information; skin spatial attention uses image gradients to enhance boundaries and local texture.

Main Results:

• AMCANet significantly outperformed existing baseline models in few-shot experiments on HAM10000 and PAD-UFES-20 datasets.
• The model demonstrated strong generalization capabilities and effective feature extraction.
• Qualitative analyses confirmed the model's ability to focus on relevant lesion regions.

Conclusions:

• AMCANet shows effectiveness in classifying skin lesions, particularly with limited sample data.
• The proposed network offers a promising direction for advancing computer-aided diagnosis in dermatology.
• The adaptive and attention-based approach enhances robustness against common challenges in skin lesion image analysis.