Burn-area estimation drops the scanner: frontal and dorsal masks alone beat the direct pixel-area ratio

Total body surface area (TBSA), the burned fraction of the whole skin surface, can now be regressed directly from frontal and dorsal whole-body masks plus burn-region masks, without a 3D scanner or a calibrated reference object. Clinical practice has relied on the rule of nines, the palm method, and Lund-Browder charts, which depend on clinician experience and vary between raters; existing 2D vision methods convert absolute area through a reference object and suffer projection errors from body curvature, while 3D methods are limited by equipment cost and acquisition time. On the MassHumanBurn test set of 8,000 burn instances and 128,000 viewpoint pairs, BurnAreaNet's mean absolute error was lower than the direct pixel-area ratio baseline, with a coefficient of determination of 0.9993 against 0.9912 and three-class accuracy of 0.984 against 0.959.

The model encodes body morphology and burn distribution separately, then uses the difference between the two feature vectors to infer the burned proportion. Two ResNet18 backbones with identical structure but non-shared weights process the whole-body masks and the burn-region masks, and their concatenated features are subtracted to form a differential feature vector. A multilayer perceptron turns that differential vector into a single TBSA float; on the CPU of an Intel N150, a 15-watt low-power processor, one batch of size one takes under 300 milliseconds.

When masks come from automatic segmentation, the method degrades less than the direct area ratio. On SAM_Preprocessed_MHB the masks were produced by Segment Anything Model 2 from box prompts derived from ground-truth connected components, so the experiment isolates mask-generation error and excludes prompt-localization error. Moving from ground-truth masks to SAM2 masks raised BurnAreaNet's mean absolute error by 0.13 percentage points and its mean relative error by 0.55 points, while the baseline rose by 0.49 and 2.42 points respectively.

The method was evaluated on synthetic data and large-scale rendered views, with inputs limited to frontal and dorsal whole-body and burn-region binary masks, and it is aimed at researchers and clinical engineers who want objective burn-area estimates from ordinary 2D imagery. The authors propose incorporating more diverse human model sources, more controllable burn synthesis, and support for additional viewpoints to cover occluded surface regions, alongside further validation on real patient data.

The authors note that the synthetic human models exclude extreme body types such as severely underweight or morbidly obese individuals, and that hair, limb amputations, and other anatomical variations were not simulated, so performance on real patient populations remains to be validated. SAM2 is a general-purpose segmentation model not designed for burn wounds and has limited effectiveness on real burn photographs, so practical deployment would require a segmentation model tailored to burn delineation. With only frontal and dorsal views, estimation becomes impossible in principle when the burn lies predominantly in regions invisible from both. Among the four high-resolution scan models, the one with BMI 39.1 performed below the baseline, which the authors say is consistent with underrepresentation in the training distribution, though a single model cannot establish causality.