Wound Segmentation with U-Net Using a Dual Attention Mechanism and Transfer Learning.

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning for dermatology

Background:

• Accurate wound segmentation is vital for diagnosing and treating skin conditions using image analysis.
• Existing methods may lack precision in identifying complex wound boundaries.
• Deep learning offers potential for automated and accurate wound segmentation.

Purpose of the Study:

• To introduce a novel dual attention U-Net model for precise wound segmentation.
• To evaluate the model's performance on various wound types, including diabetic foot ulcers, acute, and chronic wounds.
• To compare the proposed model against state-of-the-art segmentation techniques.

Main Methods:

• Development of a dual attention U-Net architecture integrating VGG16 and U-Net.
• Incorporation of dual attention mechanisms to focus on critical wound regions.
• Training and fine-tuning the model on diverse wound image datasets (diabetic foot ulcers, acute, chronic wounds).

Main Results:

• The dual attention U-Net model achieved high performance metrics.
• Achieved a Dice coefficient of 94.1% and an Intersection over Union (IoU) of 89.3% on the test set.
• Demonstrated superior performance compared to other state-of-the-art models.

Conclusions:

• The proposed dual attention U-Net model offers a robust and effective solution for precise wound segmentation.
• The model exhibits strong generalization capabilities across different wound types.
• This approach has significant potential for improving diagnostic and treatment accuracy in wound care.