Objective：
Accurate assessment of the percentage of total body surface area (% TBSA) burned is crucial in the management of burn injuries. It is difficult to estimate a size of an irregular shape by inspection. Many articles reported the discrepancy of estimating % TBSA burned by different doctors. We proposed a system of deep learning models to perform burn wound segmentation and deep burn classification.


Material and Methods：
There were three deep learning (DL) models for three tasks. Model one (M1) was responsible for total burn area segmentation. Model two (M2) was applied for deep burn area segmentation. Model three (M3) is used for palm segmentation. By giving a distance of taking photos, the % TBSA burned could be calculated by the results of M1 and M3. The percentage of the area of a deep burn can be acquired by the results of M1 and M2. Several powerful DL models (U-Net, RefineNet, PsPnet, deeplabV3+, FPN, Mask R-CNN) with different encoders (ResNet50, ResNet101) had been trained and tested. The images and data of burn victims in the clinic and ER were input to validate the performance of our system.

Results：
Total 4991 images of early burn wounds were labeled and used to train M1, while 1565 images with deep burn were labeled and used to train M2. Total 1050 images of palms from near-equal female and male populations were labeled and used to train M3. DeeplabV3+ had the best performance as M1 and M2, whereas Mask R-CNN had the best results as M3. From November 2020 to October 2021, our web-based system was applied to assess 412 burn wounds from 128 patients. Compared with grand truth labeled by plastic surgeons, it achieved satisfying results with precision:0.8804, recall: 0.8089, F1 score: 0.8117, accuracy: 0.9735, and loss: 0.1880.

Conclusions：
In the era of precision medicine, the %TBSA burned can be automatically diagnosed by DL models on a pixel-to-pixel scale. Artificial intelligence provides consistent, accurate, and rapid assessments of burn wounds.