AI-Based Wound Measurement in a Telemedicine-Style Workflow: Agreement With In-Person Reference Measurements Compared With Individual Physicians and Expert Consensus.

Authors

Tassilo Dege, Vanessa Borst, Luisa Deutzmann, Jan Hendrik Maiwald, Samuel Kounev, Astrid Schmieder

Published in

International wound journal. Volume 23. Issue 10. Pages e71043.

Abstract

Accurate wound size measurement is essential for monitoring healing, yet visual estimates are variable-especially in telemedicine-style assessments. To compare AI-based wound size estimates with physician image-based estimates using in-person measurements as ground truth (GT). In a prospective dataset of bedside wound photographs (Dec 2024-Apr 2025), GT height and width were measured in person by two experienced raters using a calibrated ruler. A set of ArUco markers placed adjacent to the wound at the same skin level enabled image-based retrieval of real-world size information. Five physicians estimated height and width from images in a telemedicine-style workflow. Two deep-learning segmentation architectures (TransNeXt-Tiny [TransNeXt] and VWFormerConvNeXtS [VW-Conv]) generated automated size estimates. Agreement and error were quantified using ICC(2,1), Pearson's r, MAE, bias and Bland-Altman limits of agreement. In addition, pixel-level segmentation performance of both models was quantified against expert-annotated reference masks using Intersection-over-Union (IoU), the Dice coefficient, precision and recall, reported as macro-averaged (per-image) and micro-averaged (pixel-pooled) values. AI achieved excellent agreement for height (TransNeXt ICC = 0.96, MAE = 0.82 cm; VW-Conv ICC = 0.95, MAE = 0.68 cm) with minimal bias. For width, AI showed moderate-to-strong agreement with a tendency to underestimation (TransNeXt ICC = 0.86, MAE = 0.69 cm, bias = -0.42 cm; VW-Conv ICC = 0.77, MAE = 0.74 cm, bias = -0.64 cm). Individual physician performance was heterogeneous, while a human consensus improved agreement (height ICC = 0.94, MAE = 0.82 cm; width ICC = 0.91, MAE = 0.65 cm). At the pixel level, TransNeXt reached a macro-averaged Dice of 0.886 (IoU 0.805) and VW-Conv a macro-averaged Dice of 0.818 (IoU 0.733), with TransNeXt showing the more consistent per-image performance. AI-based wound measurement was more consistent than single physician estimates and reached agreement comparable to, but not exceeding, a consensus of experts. Within the constraints of this single-centre validation study, these findings support AI-assisted measurement as an objective aid for standardised wound monitoring in telemedicine and clinical research; confirmation in larger, multicentre and prospectively recruited cohorts is required.

PMID:
42885306
Bibliographic data and abstract were imported from PubMed on 16 Oct 2026.

Read full publication at:
Please sign in to see all details.

Stats

1-terrible, 9-excellent. How would you rate this publication? Sign in in to submit your rating.

Post a comment

You need to be signed in to post comments. You can sign in here.

Comments

There are no comments yet.