Evaluation of Artificial Intelligence-Based Wound Assessment Service

Introduction: Accurate assessments of wound area and granulation tissue are important for tracking wound healing outcomes in response to therapies. However, many providers rely on inaccurate ruler-based measurements of wound length and width, simplifying complex wound geometries to a bounding rectangle. Manually tracing the wound improves accuracy but is too time-consuming for a busy wound clinic and varies between tracers. Further, percentage of healthy granulation tissue correlates with healing but is typically visually estimated by providers, which is also inaccurate and highly variable. Algorithms that accurately quantify wound area and granulation tissue stand to improve the consistency and efficiency of wound measurement, allowing better tracking of patient outcomes and enhancing clinical decisions based on wound data. The purpose of this study was to compare wound area and granulation tissue measurement by an artificial intelligence (AI)-based software to manual measurement performed by wound care clinicians.

Methods: In this IRB-approved retrospective study, two wound fellows and an AI-based wound assessment service independently traced wound area (n=110) and granulation tissue area (n=25) on a set of 110 wound images taken for routine care at a tertiary wound center. Differences between pairs of tracings were determined by comparing a “reference” manual trace (either human) to a “test” trace (AI or the other human) and calculating relative error and false negative/positive areas. Three blinded attending physicians manually assessed granulation area and qualitatively assessed tracing accuracy.

Results: There were no significant differences in error measures between human-human and AI-human comparisons for both wound and granulation area. Visual estimates of granulation area varied widely across attendings, and accuracy was qualitatively similar between humans and AI.

Conclusion: Similarity in error measures between AI-human and human-human comparisons suggests performance on par with human tracers. Standardization of wound and granulation area using algorithms stands to improve data quality in wound care.