Improving Clinical Generalization of Pressure Ulcer Stage Classification Through Saliency-Guided Data Augmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Dermatology

Background:

• Medical imaging datasets for specific conditions like pressure ulcers are often limited.
• Variations in imaging conditions (distance, lighting, viewpoint) hinder accurate clinical classification.
• Developing robust AI models for pressure ulcer staging is crucial for effective patient care.

Purpose of the Study:

• To improve the generalization capability of AI models for pressure ulcer stage classification.
• To address challenges posed by data scarcity and variability in clinical medical imaging.
• To enhance the clinical usability of AI-powered diagnostic tools.

Main Methods:

• Developed a YOLOv7-based model for pressure ulcer stage classification.
• Employed a two-phase training strategy incorporating saliency-guided images.
• Utilized clinically plausible noise augmentation, including healing areas and white keratin.

Main Results:

• Achieved an accuracy increase from 75% to 89% on newly acquired hospital images.
• Demonstrated stable and reproducible performance with five-fold cross-validation (mAP@0.5: 86.20% ± 2.28%).
• Exceeded prior reported performance benchmarks for pressure ulcer staging models in clinical settings.

Conclusions:

• Curriculum learning combined with noise-enriched augmentation improves model generalization in clinical settings.
• Clinically informed data augmentation is essential for enhancing AI model performance in medical imaging.
• The proposed approach offers a practical solution for improving AI usability in data-limited medical environments.