A multimodal skin lesion classification through cross-attention fusion and collaborative edge computing.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Skin cancer diagnosis relies on accurate and timely identification.
• Deep learning computer-aided diagnosis (CAD) systems show promise but face challenges like low accuracy, high latency, and data privacy concerns.

Purpose of the Study:

• To develop a unified solution for skin lesion classification addressing limitations of current CAD systems.
• To enhance diagnostic accuracy, reduce network latency, and improve data privacy in skin cancer diagnosis.

Main Methods:

• A multimodal deep learning model integrating dermoscopic images and patient metadata using cross-attention feature fusion.
• A collaborative inference scheme distributing computations across Internet of Things (IoT) and edge devices for local processing.

Main Results:

• Achieved 95.73% classification accuracy on the HAM10000 dataset, outperforming existing methods.
• Demonstrated significant efficiency improvements with latency speedups of up to 20% (device-only) and 47% (edge-only).
• Validated the model's effectiveness and generalizability across multiple benchmark datasets.

Conclusions:

• The proposed unified solution effectively enhances skin lesion classification accuracy and efficiency.
• Collaborative inference on edge devices significantly improves data privacy and reduces latency in medical AI applications.
• This approach offers a robust and scalable framework for advanced skin cancer diagnostic support systems.