Facial wrinkle segmentation using weighted deep supervision and semi-automatic labeling.

Area of Science:

• Biomedical Imaging
• Computer Vision
• Dermatology

Background:

• Facial wrinkles are key indicators of aging, but their segmentation is challenging due to thinness and small image proportion.
• Deep learning methods show promise but require further performance improvements for accurate wrinkle segmentation.

Purpose of the Study:

• To develop a novel loss function for improved wrinkle segmentation by incorporating wrinkle thickness.
• To introduce a weighted deep supervision approach for more accurate training loss computation.

Main Methods:

• Proposed a weighted deep supervision method using weighted wrinkle maps generated from ground truth.
• Utilized a U-Net architecture and evaluated the method on a skin analysis device dataset with semi-automatic labeling.
• Assessed scalability by applying the method to retinal vessel segmentation.

Main Results:

• The weighted deep supervision method achieved higher Jaccard Similarity Index (JSI) for wrinkle segmentation compared to conventional methods.
• Semi-automatic labeling demonstrated more consistent wrinkle labels than human labeling.
• The method showed superior performance in retinal vessel segmentation, indicating broad applicability.

Conclusions:

• The proposed weighted deep supervision method significantly enhances wrinkle segmentation accuracy.
• Semi-automatic labeling offers a reliable alternative to human labeling for biomedical image analysis.
• The method is highly performant, scalable across biomedical domains, and compatible with U-Net architectures.