Abstract / Summary
Accurate estimation of total body surface area (TBSA) is critical for burn diagnosis, triage, and prognosis. It represents the burned area as a proportion of the total body surface area, and is usually expressed as a percentage value. Traditional methods are subjective, while existing computer-aided approaches require specialized equipment or suffer from projection errors. To address these limitations, we constructed MassHumanBurn, a synthetic dataset comprising 5,000 diverse human models with simulated burns across eight TBSA levels, yielding 320,000 rendered views with paired body and burn masks. We propose BurnAreaNet, a neural network that estimates TBSA from frontal and dorsal whole-body and burn region masks. On ground-truth masks, BurnAreaNet achieved a descriptive mean absolute error (MAE) of \(0.55\% \pm 0.49\%\) , substantially outperforming the direct area ratio baseline ( \(1.99\% \pm 1.66\%\) ). To evaluate the impact of segmentation errors, we further constructed SAM_Preprocessed_MHB, where masks were generated using Segment Anything Model 2 (SAM2). When masks were generated by SAM2 using box prompts derived from ground-truth masks, BurnAreaNet achieved an MAE of \(0.68\% \pm 0.70\%\) , while the baseline method yielded an MAE of \(2.48\% \pm 3.52\%\) . In a preliminary evaluation on four high-resolution human scan models (HSR_Burn dataset), BurnAreaNet outperformed the baseline on three of four models, while performance was lower for one high-BMI model (BMI 39.1) that may be underrepresented in the training distribution. These results suggest that BurnAreaNet represents a promising step toward rapid, objective, and device-independent TBSA estimation. Future work should incorporate more diverse human model sources, more controllable burn synthesis methods, and support for additional viewpoints to address occluded surface regions, while further validation on real patient data remains essential.