Back to 2026 Abstracts
Artificial Intelligence Modeling for Pediatric Burn Measurement: A Feasibility Study for TBSA Calculation
Meagan Rosenberg1, Jennifer Murphy2, Emily Ibrahimov, Victoria Pepper3, Gregory Banever3, David Tashjian3, Kevin Moriarty3, Mark Roberson4, Michael Tirabassi3 1 Baystate Medical Center, Department of General Surgery, Springfield, MA 2 University of Massachusetts Chan Medical School, Worcester, MA 3 Baystate Medical Center, Division of Pediatric Surgery, Springfield, MA 4 Goldfinch Sensors and Technology, Cary, NC
Background: Burn percentage total body surface area(TBSA) estimates are often imprecise with clinically impactful disparities between care settings. This study evaluates the feasibility of segmentation modeling as a novel tool to measure burns in pediatric patients. Methods: Laerdal Medical’s SimJunior® was used as a model with felt patches as simulated burns. Forty-one images were obtained with the burns in unique configurations. Images were annotated using Smart Select with human refinement of the outlines. Augmentation with flips and 15° lateral skews generated 124 images. Ten were used for testing, 10 for validation, and 10 for training. Roboflow’s YOLO26 instance segmentation was utilized with mean average precision(mAP) and loss metrics to monitor performance. Results: Three hundred-five epochs were completed. After initial variability, mAP plateaued around 98%. The mAP@50:95 stabilized around 95%(Fig 1). Object and class loss both approached zero at the end of training. Box loss showed a transient rise before stabilizing around 0.20. The early, transient rise in object loss is reflected in the model’s early misses of hand, face, and foot burns on test images. With manual correction, these false negatives decreased in frequency. There were no false positives on testing images. Conclusion: This pilot instance segmentation model, although limited by small sample size, achieved a high mAP@50:95 of 95%. Precise burn border identification is a fundamental step towards using AI-assistance to calculate percentage TBSA.
Back to 2026 Abstracts
