A novel RSCA-YOLOv8s network for automatic diagnosis and graduation in pressure injury

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate pressure injury (PI) diagnosis is crucial for patient care.
• Current diagnostic methods can lack objectivity and consistency.
• Advanced AI models show promise for improving diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate an improved YOLOv8s model for objective and accurate PI recognition and staging.
• To enhance feature extraction and multi-scale recognition capabilities for PI detection.
• To provide a reliable AI-assisted tool for clinical PI diagnosis.

Main Methods:

• An improved YOLOv8s model was developed incorporating spatial and channel attention mechanisms and a multi-scale fusion module.
• A dataset of 366 PI images was collected from two hospitals and split into training (80%) and validation (20%) sets.
• The improved YOLOv8s was trained and compared against YOLOv5, TPH-YOLO, YOLOv7, and Swin Transformer using precision, recall, F1-score, mAP50, and mAP50:95.

Main Results:

• The improved YOLOv8s achieved a mAP50 of 92.0% and precision (P) of 86.7%, outperforming other evaluated models.
• Compared to the original YOLOv8s, the improved version showed a 6.5% increase in precision, 14% in mAP50, and 14.9% in mAP50:95.
• Recognition accuracy for PI stages 1-4 was 89.3%, 84.3%, 73.2%, and 100%, respectively.

Conclusions:

• The improved YOLOv8s model demonstrates significant effectiveness in recognizing pressure injuries across different stages.
• The proposed model offers an objective and reliable auxiliary tool for clinical PI diagnosis.
• The integration of attention mechanisms and multi-scale fusion enhances the model's diagnostic performance.