A Multicenter Study of the Evaluation of an Artificial Intelligence-based Automated Software for Wound Assessment

Abstract Body: Introduction: Wound area and healing progression is measured by multiplying the greatest length and width across the diameter regardless of wound shape. This ruler-based measurement has been reported to result in ≥40% inaccuracy. Manual tracing of the wound area (planimetry) improves this accuracy by addressing the wound shape, but is time consuming in a fast-paced and high volume wound clinic. Both methods, however, are subjective and provider dependent. Computer-based automated planimetry can potentially provide accurate, fast, and unbiased measurements. This study compared wound planimetry by an artificial intelligence (AI)-based automated software to manual planimetry performed by an expert panel of wound care providers. Methods: This was an IRB-approved, retrospective multi-center study with two tertiary care wound healing centers that included 215 wound photographs from May 2018 to August 2019. All wound types and stages of healing were included. The wound area to be traced was defined as the largest, continuous non-epithelialized area of the wound including any satellite lesion areas within two centimeters. For each photo, the wound was traced manually by two non-blinded wound clinicians using ImageJ software and also automatically using AI software. Using each manual tracing as a standard, the relative errors in area were calculated for each comparison (AI-manual and manual-manual) and statistically evaluated by repeated measures ANOVA. Three blinded expert physicians from both centers were shown the three tracings and asked to determine the most accurate image and whether each met the wound area definition. Results: At the first site, median relative errors between AI and each human tracing were 12.4% and 10.2%; between the two human tracings were 8.0% and 8.4%. Repeated measures ANOVA did not detect significant differences in mean relative error across comparisons between AI-manual or manual-manual traces.  On average, traces were classified as meeting the wound area definition for 66.8% (manual), 57.1% (manual), and 51.5% (AI) of photos. Reviewers chose human-annotated images as most accurate for 81.2% of the photos. Result analysis from the second site are pending.  Conclusion: AI can improve efficiency and decrease wound measurement inaccuracies. Despite human expertise, future work with planimetry to standardize wound measurement is warranted.