BurnCare: A deep learning framework for burn injury assessment and TBSA estimation.

BurnCare estimates burn %TBSA from a single smartphone photo, achieving strong agreement with clinicians (r = 0.746, MAE = 6.29 percentage points; n = 16).

• Why it matters: Resource-limited settings in India lack dedicated burn centres, so non-specialists often estimate %TBSA with known inter-observer variability that affects resuscitation and referral decisions.
• What they did: BurnCare is a two-stage deep learning pipeline: ResNet50-U-Net background removal and DenseNet169-U-Net burn segmentation, trained on a prospectively collected multi-device dataset from a tertiary burn centre in India. It outputs %TBSA with or without scale calibration and compares results to blinded clinician estimates using Pearson correlation, MAE, RMSE, and ICC.
• The result: BurnCare’s automatic %TBSA correlates strongly with the clinicians’ mean (r = 0.746, p < 0.001; MAE = 6.29; RMSE = 8.63), improving after manual mask correction (r = 0.853; MAE = 4.32; RMSE = 6.37), with clinician-only ICC = 0.728.