A multi-head YOLOv12 with self-supervised pretraining for urinary sediment particle detection

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Computer Vision

Background:

• Manual urine sediment analysis is crucial but suffers from subjectivity and time constraints.
• Existing automated methods lack the granularity for comprehensive particle classification.
• Developing an objective and efficient automated system is vital for clinical practice.

Purpose of the Study:

• To develop a novel deep learning method for automated and reliable urine sediment analysis.
• To enable simultaneous, fine-grained classification of diverse urine sediment particles.
• To overcome limitations of manual analysis and prior automated techniques.

Main Methods:

• Utilized a multi-head YOLOv12 architecture with self-supervised pretraining.
• Implemented Slicing Aided Hyper Inference (SAHI) for advanced object detection.
• Created a large-scale dataset (OpenUrine) with 790 labeled and 5640 unlabeled images across 39 categories.

Main Results:

• Achieved 76.59% precision and 64.15% mean Average Precision (mAP) on a 39-class dataset.
• Demonstrated competitive detection accuracy, particularly for small and low-contrast objects.
• The six specialized detection heads enabled comprehensive particle classification.

Conclusions:

• The proposed deep learning method offers a significant advancement in automated urine sediment analysis.
• This AI-driven approach enhances diagnostic capabilities for renal and urinary disorders.
• The model's performance indicates its potential for clinical application and improved patient outcomes.