Novel two-stage deep learning framework for automated pressure injury classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Accurate pressure injury (PI) staging is crucial for effective treatment and patient outcomes.
• Current PI staging relies on subjective clinical assessment, leading to variability.
• Automating PI staging from clinical images can enhance diagnostic consistency and efficiency.

Purpose of the Study:

• To develop and validate a two-stage artificial intelligence (AI) framework for automatic pressure injury (PI) staging directly from raw clinical images.
• To improve diagnostic accuracy and reduce subjectivity in PI staging.
• To support consistent PI staging within clinical workflows.

Main Methods:

• A retrospective study utilized 1807 PI images with senior nurse annotations.
• A two-stage deep learning approach was employed: YOLOv9 for lesion detection and DenseNet161 for staging.
• Performance was evaluated using metrics including accuracy, sensitivity, specificity, F1 score, AUC, and mAP@0.5.

Main Results:

• The YOLOv9 object detection model achieved a mean average precision (mAP@0.5) of 0.796.
• The staging model demonstrated an overall accuracy of 0.775, sensitivity of 0.775, and specificity of 0.955 on an independent test set.
• The framework showed performance comparable to or exceeding previous methods, with improved clinical interpretability.

Conclusions:

• The proposed AI framework effectively performs PI staging from clinical images, demonstrating robust performance.
• The system offers standardized and reproducible PI staging, with potential for integration into nursing workflows.
• Further development is needed to address challenges like intra-wound heterogeneity and image quality variability.