Semi-Supervised Facial Acne Segmentation Using Bidirectional Copy-Paste.

Area of Science:

• Dermatology
• Computer Vision
• Artificial Intelligence

Background:

• Facial acne is a common skin condition requiring early detection to prevent worsening.
• Automated detection using deep learning shows promise but is hindered by the difficulty of acquiring sufficient labeled acne training data.
• Existing semi-supervised methods for medical image segmentation often struggle with performance, particularly in the labeled data training component.

Purpose of the Study:

• To propose a novel deep learning model for facial acne segmentation.
• To address the challenge of limited labeled data in acne detection using a semi-supervised approach.
• To enhance the performance of semi-supervised learning for facial acne segmentation.

Main Methods:

• Developed a novel deep learning model for facial acne segmentation.
• Utilized a bidirectional copy-paste semi-supervised learning technique to synthesize training images by exchanging foreground and background components between labeled and unlabeled datasets.
• Implemented a new framework to directly compute training loss on labeled images, improving performance over previous methods.

Main Results:

• The proposed method achieved a Dice score of 0.5205 in experiments using only 3% labeled data.
• Demonstrated an improvement of 0.0151 to 0.0473 in Dice score compared to existing semi-supervised learning methods.
• Showcased superior performance in facial acne segmentation, particularly under conditions of extreme data scarcity.

Conclusions:

• The novel semi-supervised learning approach significantly improves facial acne segmentation performance.
• The bidirectional copy-paste method offers an effective solution for training deep learning models with limited labeled dermatological data.
• This research provides a valuable new direction for automated acne analysis and dermatological diagnostics.