SA-RPN: A Spacial Aware Region Proposal Network for Acne Detection

Area of Science:

• Dermatology
• Computer Vision
• Medical Imaging Analysis

Background:

• Automated detection of acne vulgaris lesions is crucial for diagnosis and treatment.
• Challenges in acne detection include blurry boundaries and small lesion sizes, hindering traditional methods.
• Existing datasets lack the high resolution and detailed annotations needed for comprehensive acne detection studies.

Purpose of the Study:

• To introduce AcneSCU, a novel, high-resolution benchmark dataset for acne lesion detection.
• To propose a Spatial Aware Region Proposal Network (SA-RPN) to enhance two-stage object detection for acne.
• To improve the accuracy and robustness of automated acne lesion detection systems.

Main Methods:

• Construction of the AcneSCU dataset with 276 high-resolution facial images and 31,777 instance-level annotations.
• Development of the Spatial Aware Region Proposal Network (SA-RPN) with disentangled representation learning and Normalized Wasserstein Distance prediction.
• Integration of SA-RPN as a plug-and-play module into standard two-stage object detection frameworks.

Main Results:

• The proposed SA-RPN significantly improves proposal quality for hard-to-detect acne samples.
• Experiments on AcneSCU and ACNE04 datasets demonstrate SA-RPN's consistent outperformance over state-of-the-art methods.
• The disentangled heads and IoU prediction enhance the correlation between classification scores and localization accuracy.

Conclusions:

• The AcneSCU dataset provides a valuable resource for advancing research in automated acne detection.
• SA-RPN effectively addresses the challenges of detecting small and blurry acne lesions.
• The proposed method offers a significant advancement in the field of dermatological image analysis and computer-aided diagnosis.