Innovative Two-Phase Deep Learning System for Automated Classification of Pressure Injuries
-
By
-
Ting-Yu Lai
-
Yi-Jiun Chou
-
Chun-You Liu
-
Chien-Wei Chen
-
Ching-Ting Lin
-
Wei-Chun Wang
-
Yimin Hsu
-
Ming-Li Hsieh
-
Shih-Sheng Chang
-
March 27, 2026
Objective:
To develop an AI framework for automatic pressure injury (PI) staging from raw clinical images without manual lesion localisation.
Approach:
- Study Design: Retrospective study using 1807 PI images from China Medical University Hospital collected between 2020 and 2024.
- Model Development: Utilized YOLOv9 for lesion detection and DenseNet161 for staging.
- Performance Metrics: Evaluated using accuracy, sensitivity, specificity, F1 score, AUC, and mAP@0.5.
Key Findings:
- YOLOv9 achieved an mAP@0.5 of 0.796.
- The staging model showed an overall accuracy of 0.775, sensitivity of 0.775, specificity of 0.955, precision of 0.779, and F1 score of 0.775.
Interpretation:
The two-stage AI framework demonstrated performance comparable to or exceeding previous approaches.
Limitations:
- Challenges in handling intra-wound heterogeneity and subtle early-stage boundaries.
- Variability in image quality may affect model performance.
Conclusion:
This two-stage AI system provides standardized, reproducible PI staging from clinical images.
