LCSD-Net: a light-weight cross-attention-based semantic dual transformer for domain generalization in melanoma...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models for melanoma detection often rely on irrelevant visual artifacts, leading to poor generalization in clinical settings.
• This bias compromises the reliability of automated skin lesion analysis.
• Developing robust and generalizable AI for melanoma detection remains a significant challenge.

Purpose of the Study:

• To propose a lightweight cross-attention-based semantic dual (LCSD) transformer model for reliable skin lesion detection.
• To address the limitations of existing models in generalizing to diverse clinical environments.
• To improve the accuracy and robustness of melanoma detection while reducing computational complexity.

Main Methods:

• The LCSD model utilizes global-level semantic information extraction and feature normalization.
• It employs semantic queries with multihead attention to refine global features and enhance domain generalization.
• Cross-attention between feature maps and semantic queries provides generalized global context encoding, reducing computational complexity from O(n^2d) to O(nmd + m^2d).

Main Results:

• The LCSD model achieved high classification accuracies on challenging datasets: 82.82% (Derm7pt-Dermoscopic), 72.95% (Derm7pt-Clinical), and 86.21% (PAD-UFES-20).
• Demonstrated superior performance compared to conventional transformer-based models.
• Showcased improved robustness and reduced computational cost.

Conclusions:

• The LCSD model effectively mitigates bias from irrelevant visual characteristics, enhancing domain generalization for melanoma detection.
• Its lightweight design facilitates deployment in mobile applications, offering a reliable and efficient solution.
• The model demonstrates significant potential for real-world melanoma detection systems across diverse clinical scenarios.