Skin Lesion Area Segmentation Using Attention Squeeze U-Net for Embedded Devices.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Melanoma is a dangerous skin cancer where early diagnosis is vital.
• Tracking lesion changes in dermoscopic images aids malignant lesion detection.
• Deep learning offers potential for automated analysis of skin lesions.

Purpose of the Study:

• To develop a deep learning architecture for skin lesion segmentation on embedded devices.
• To enhance patient empowerment by enabling local analysis on smartphones.
• To facilitate lesion history tracking, reduce hospital visits, and protect user privacy.

Main Methods:

• A novel deep learning architecture, Attention Squeeze U-Net, was designed for efficient skin lesion segmentation.
• The model is optimized for deployment on resource-constrained embedded devices and smartphones.
• Publicly available dermoscopic image datasets were used for quantitative evaluation.

Main Results:

• The Attention Squeeze U-Net achieved effective skin lesion segmentation.
• Good segmentation performance was demonstrated even with a compact model size.
• The architecture is suitable for deployment on low-cost embedded devices.

Conclusions:

• Deep learning models like Attention Squeeze U-Net can be effectively deployed on smartphones for skin lesion analysis.
• Local processing of dermoscopic images enhances patient privacy and accessibility.
• This approach supports early melanoma detection and patient self-monitoring.